new

.gitignore

@@ -10,3 +10,5 @@ wheels/
 .venv
 .hypothesis
 samples
+*.jpg
+*.parquet

app/camera/__init__.py

@@ -1,6 +1,7 @@
 from collections import OrderedDict, defaultdict
 from dataclasses import dataclass
 from datetime import datetime
+import string
 from typing import Any, TypeAlias, TypedDict, Optional, Sequence
 
 from beartype import beartype
@@ -522,10 +523,14 @@ def to_homogeneous(points: Num[Array, "N 2"] | Num[Array, "N 3"]) -> Num[Array,
         raise ValueError(f"Invalid shape for points: {points.shape}")
 
 
+import awkward as ak
+
+
 @jaxtyped(typechecker=beartype)
 def point_line_distance(
     points: Num[Array, "N 3"] | Num[Array, "N 2"],
     line: Num[Array, "N 3"],
+    description: str,
     eps: float = 1e-9,
 ):
     """
@@ -544,6 +549,12 @@ def point_line_distance(
     """
     numerator = abs(line[:, 0] * points[:, 0] + line[:, 1] * points[:, 1] + line[:, 2])
     denominator = jnp.sqrt(line[:, 0] * line[:, 0] + line[:, 1] * line[:, 1])
+    # line_data = {"a": line[:, 0], "b": line[:, 1], "c": line[:, 2]}
+    # line_x_y = {"x": points[:, 0], "y": points[:, 1]}
+    # ak.to_parquet(
+    #     line_data, f"/home/admin/Code/CVTH3PE/line_a_b_c_{description}.parquet"
+    # )
+    # ak.to_parquet(line_x_y, f"/home/admin/Code/CVTH3PE/line_x_y_{description}.parquet")
     return numerator / (denominator + eps)
 
 
@@ -571,7 +582,7 @@ def left_to_right_epipolar_distance(
     """
     F_t = fundamental_matrix.transpose()
     line1_in_2 = jnp.matmul(left, F_t)
-    return point_line_distance(right, line1_in_2)
+    return point_line_distance(right, line1_in_2, "left_to_right")
 
 
 @jaxtyped(typechecker=beartype)
@@ -597,7 +608,7 @@ def right_to_left_epipolar_distance(
         $$x^{\\prime T}Fx = 0$$
     """
     line2_in_1 = jnp.matmul(right, fundamental_matrix)
-    return point_line_distance(left, line2_in_1)
+    return point_line_distance(left, line2_in_1, "right_to_left")
 
 
 def distance_between_epipolar_lines(

app/tracking/__init__.py

@@ -1,14 +1,20 @@
+import warnings
+import weakref
+from collections import deque
 from dataclasses import dataclass
-from datetime import datetime
+from datetime import datetime, timedelta
+from itertools import chain
 from typing import (
     Any,
     Callable,
     Generator,
     Optional,
+    Protocol,
     Sequence,
     TypeAlias,
     TypedDict,
     TypeVar,
+    Union,
     cast,
     overload,
 )
@@ -18,18 +24,428 @@ from beartype import beartype
 from beartype.typing import Mapping, Sequence
 from jax import Array
 from jaxtyping import Array, Float, Int, jaxtyped
-from pyrsistent import PVector
+from pyrsistent import PVector, v, PRecord, PMap
 
-from app.camera import Detection
+from app.camera import Detection, CameraID
+
+TrackingID: TypeAlias = int
+
+
+class TrackingPrediction(TypedDict):
+    velocity: Optional[Float[Array, "J 3"]]
+    keypoints: Float[Array, "J 3"]
+
+
+class GenericVelocityFilter(Protocol):
+    """
+    a filter interface for tracking velocity estimation
+    """
+
+    def predict(self, timestamp: datetime) -> TrackingPrediction:
+        """
+        predict the velocity and the keypoints location
+
+        Args:
+            timestamp: timestamp of the prediction
+
+        Returns:
+            velocity: velocity of the tracking
+            keypoints: keypoints of the tracking
+        """
+        ...  # pylint: disable=unnecessary-ellipsis
+
+    def update(self, keypoints: Float[Array, "J 3"], timestamp: datetime) -> None:
+        """
+        update the filter state with new measurements
+
+        Args:
+            keypoints: new measurements
+            timestamp: timestamp of the update
+        """
+        ...  # pylint: disable=unnecessary-ellipsis
+
+    def get(self) -> TrackingPrediction:
+        """
+        get the current state of the filter state
+
+        Returns:
+            velocity: velocity of the tracking
+            keypoints: keypoints of the tracking
+        """
+        ...  # pylint: disable=unnecessary-ellipsis
+
+
+class DummyVelocityFilter(GenericVelocityFilter):
+    """
+    a dummy velocity filter that does nothing
+    """
+
+    _keypoints_shape: tuple[int, ...]
+
+    def __init__(self, keypoints: Float[Array, "J 3"]):
+        self._keypoints_shape = keypoints.shape
+
+    def predict(self, timestamp: datetime) -> TrackingPrediction:
+        return TrackingPrediction(
+            velocity=None,
+            keypoints=jnp.zeros(self._keypoints_shape),
+        )
+
+    def update(self, keypoints: Float[Array, "J 3"], timestamp: datetime) -> None: ...
+
+    def get(self) -> TrackingPrediction:
+        return TrackingPrediction(
+            velocity=None,
+            keypoints=jnp.zeros(self._keypoints_shape),
+        )
+
+
+class LastDifferenceVelocityFilter(GenericVelocityFilter):
+    """
+    a naive velocity filter that uses the last difference of keypoints
+    """
+
+    _last_timestamp: datetime
+    _last_keypoints: Float[Array, "J 3"]
+    _last_velocity: Optional[Float[Array, "J 3"]] = None
+
+    def __init__(self, keypoints: Float[Array, "J 3"], timestamp: datetime):
+        self._last_keypoints = keypoints
+        self._last_timestamp = timestamp
+
+    def predict(self, timestamp: datetime) -> TrackingPrediction:
+        delta_t_s = (timestamp - self._last_timestamp).total_seconds()
+        if delta_t_s <= 0:
+            warnings.warn(
+                "delta_t={}; last={}; current={}".format(
+                    delta_t_s, self._last_timestamp, timestamp
+                )
+            )
+        if self._last_velocity is None:
+            return TrackingPrediction(
+                velocity=None,
+                keypoints=self._last_keypoints,
+            )
+        else:
+            if delta_t_s <= 0:
+                return TrackingPrediction(
+                    velocity=self._last_velocity,
+                    keypoints=self._last_keypoints,
+                )
+            return TrackingPrediction(
+                velocity=self._last_velocity,
+                keypoints=self._last_keypoints + self._last_velocity * delta_t_s,
+            )
+
+    def update(self, keypoints: Float[Array, "J 3"], timestamp: datetime) -> None:
+        delta_t_s = (timestamp - self._last_timestamp).total_seconds()
+        if delta_t_s <= 0:
+            pass
+        else:
+            self._last_timestamp = timestamp
+            self._last_velocity = (keypoints - self._last_keypoints) / delta_t_s
+        self._last_keypoints = keypoints
+
+    def get(self) -> TrackingPrediction:
+        if self._last_velocity is None:
+            return TrackingPrediction(
+                velocity=None,
+                keypoints=self._last_keypoints,
+            )
+        else:
+            return TrackingPrediction(
+                velocity=self._last_velocity,
+                keypoints=self._last_keypoints,
+            )
+
+
+class LeastMeanSquareVelocityFilter(GenericVelocityFilter):
+    """
+    a velocity filter that uses the least mean square method to estimate the velocity
+    """
+
+    _historical_3d_poses: deque[Float[Array, "J 3"]]
+    _historical_timestamps: deque[datetime]
+    _velocity: Optional[Float[Array, "J 3"]] = None
+    _max_samples: int
+
+    def __init__(
+        self,
+        historical_3d_poses: Sequence[Float[Array, "J 3"]],
+        historical_timestamps: Sequence[datetime],
+        max_samples: int = 10,
+    ):
+        assert len(historical_3d_poses) == len(historical_timestamps)
+        temp = zip(historical_3d_poses, historical_timestamps)
+        temp_sorted = sorted(temp, key=lambda x: x[1])
+        self._historical_3d_poses = deque(
+            map(lambda x: x[0], temp_sorted), maxlen=max_samples
+        )
+        self._historical_timestamps = deque(
+            map(lambda x: x[1], temp_sorted), maxlen=max_samples
+        )
+        self._max_samples = max_samples
+        if len(self._historical_3d_poses) < 2:
+            self._velocity = None
+        else:
+            self._update(
+                jnp.array(self._historical_3d_poses),
+                jnp.array(self._historical_timestamps),
+            )
+
+    def predict(self, timestamp: datetime) -> TrackingPrediction:
+        if not self._historical_3d_poses:
+            raise ValueError("No historical 3D poses available for prediction")
+
+        # use the latest historical detection
+        latest_3d_pose = self._historical_3d_poses[-1]
+        latest_timestamp = self._historical_timestamps[-1]
+
+        delta_t_s = (timestamp - latest_timestamp).total_seconds()
+
+        if self._velocity is None:
+            return TrackingPrediction(
+                velocity=None,
+                keypoints=latest_3d_pose,
+            )
+        else:
+            # Linear motion model: ẋt = xt' + Vt' · (t - t')
+            predicted_3d_pose = latest_3d_pose + self._velocity * delta_t_s
+            return TrackingPrediction(
+                velocity=self._velocity, keypoints=predicted_3d_pose
+            )
+
+    @jaxtyped(typechecker=beartype)
+    def _update(
+        self,
+        keypoints: Float[Array, "N J 3"],
+        timestamps: Float[Array, "N"],
+    ) -> None:
+        """
+        update measurements with least mean square method
+        """
+        if keypoints.shape[0] < 2:
+            raise ValueError("Not enough measurements to estimate velocity")
+
+        # Using least squares to fit a linear model for each joint and dimension
+        # X = timestamps, y = keypoints
+        # For each joint and each dimension, we solve for velocity
+
+        n_samples = timestamps.shape[0]
+        n_joints = keypoints.shape[1]
+
+        # Create design matrix for linear regression
+        # [t, 1] for each timestamp
+        X = jnp.column_stack([timestamps, jnp.ones(n_samples)])
+
+        # Reshape keypoints to solve for all joints and dimensions at once
+        # From [N, J, 3] to [N, J*3]
+        keypoints_reshaped = keypoints.reshape(n_samples, -1)
+
+        # Use JAX's lstsq to solve the least squares problem
+        # This is more numerically stable than manually computing pseudoinverse
+        coefficients, _, _, _ = jnp.linalg.lstsq(X, keypoints_reshaped, rcond=None)
+
+        # Coefficients shape is [2, J*3]
+        # First row: velocities, Second row: intercepts
+        velocities = coefficients[0].reshape(n_joints, 3)
+
+        # Update velocity
+        self._velocity = velocities
+
+    def update(self, keypoints: Float[Array, "J 3"], timestamp: datetime) -> None:
+        last_timestamp = self._historical_timestamps[-1]
+        assert last_timestamp <= timestamp
+
+        # deque would manage the maxlen automatically
+        self._historical_3d_poses.append(keypoints)
+        self._historical_timestamps.append(timestamp)
+
+        t_0 = self._historical_timestamps[0]
+        all_keypoints = jnp.array(self._historical_3d_poses)
+
+        def timestamp_to_seconds(timestamp: datetime) -> float:
+            assert t_0 <= timestamp
+            return (timestamp - t_0).total_seconds()
+
+        # timestamps relative to t_0 (the oldest detection timestamp)
+        all_timestamps = jnp.array(
+            map(timestamp_to_seconds, self._historical_timestamps)
+        )
+
+        self._update(all_keypoints, all_timestamps)
+
+    def get(self) -> TrackingPrediction:
+        if not self._historical_3d_poses:
+            raise ValueError("No historical 3D poses available")
+
+        latest_3d_pose = self._historical_3d_poses[-1]
+
+        if self._velocity is None:
+            return TrackingPrediction(velocity=None, keypoints=latest_3d_pose)
+        else:
+            return TrackingPrediction(velocity=self._velocity, keypoints=latest_3d_pose)
+
+
+class OneEuroFilter(GenericVelocityFilter):
+    """
+    Implementation of the 1€ filter (One Euro Filter) for smoothing keypoint data.
+
+    The 1€ filter is an adaptive low-pass filter that adjusts its cutoff frequency
+    based on movement speed to reduce jitter during slow movements while maintaining
+    responsiveness during fast movements.
+
+    Reference: https://cristal.univ-lille.fr/~casiez/1euro/
+    """
+
+    _x_filtered: Float[Array, "J 3"]
+    _dx_filtered: Optional[Float[Array, "J 3"]] = None
+    _last_timestamp: datetime
+    _min_cutoff: float
+    _beta: float
+    _d_cutoff: float
+
+    def __init__(
+        self,
+        keypoints: Float[Array, "J 3"],
+        timestamp: datetime,
+        min_cutoff: float = 1.0,
+        beta: float = 0.0,
+        d_cutoff: float = 1.0,
+    ):
+        """
+        Initialize the One Euro Filter.
+
+        Args:
+            keypoints: Initial keypoints positions
+            timestamp: Initial timestamp
+            min_cutoff: Minimum cutoff frequency (lower = more smoothing)
+            beta: Speed coefficient (higher = less lag during fast movements)
+            d_cutoff: Cutoff frequency for the derivative filter
+        """
+        self._last_timestamp = timestamp
+
+        # Filter parameters
+        self._min_cutoff = min_cutoff
+        self._beta = beta
+        self._d_cutoff = d_cutoff
+
+        # Filter state
+        self._x_filtered = keypoints  # Position filter state
+        self._dx_filtered = None  # Initially no velocity estimate
+
+    @overload
+    def _smoothing_factor(self, cutoff: float, dt: float) -> float: ...
+
+    @overload
+    def _smoothing_factor(
+        self, cutoff: Float[Array, "J"], dt: float
+    ) -> Float[Array, "J"]: ...
+
+    @jaxtyped(typechecker=beartype)
+    def _smoothing_factor(
+        self, cutoff: Union[float, Float[Array, "J"]], dt: float
+    ) -> Union[float, Float[Array, "J"]]:
+        """Calculate the smoothing factor for the low-pass filter."""
+        r = 2 * jnp.pi * cutoff * dt
+        return r / (r + 1)
+
+    @jaxtyped(typechecker=beartype)
+    def _exponential_smoothing(
+        self,
+        a: Union[float, Float[Array, "J"]],
+        x: Float[Array, "J 3"],
+        x_prev: Float[Array, "J 3"],
+    ) -> Float[Array, "J 3"]:
+        """Apply exponential smoothing to the input."""
+        return a * x + (1 - a) * x_prev
+
+    def predict(self, timestamp: datetime) -> TrackingPrediction:
+        """
+        Predict keypoints position at a given timestamp.
+
+        Args:
+            timestamp: Timestamp for prediction
+
+        Returns:
+            TrackingPrediction with velocity and keypoints
+        """
+        dt = (timestamp - self._last_timestamp).total_seconds()
+
+        if self._dx_filtered is None:
+            return TrackingPrediction(
+                velocity=None,
+                keypoints=self._x_filtered,
+            )
+        else:
+            predicted_keypoints = self._x_filtered + self._dx_filtered * dt
+            return TrackingPrediction(
+                velocity=self._dx_filtered,
+                keypoints=predicted_keypoints,
+            )
+
+    def update(self, keypoints: Float[Array, "J 3"], timestamp: datetime) -> None:
+        """
+        Update the filter with new measurements.
+
+        Args:
+            keypoints: New keypoint measurements
+            timestamp: Timestamp of the measurements
+        """
+        dt = (timestamp - self._last_timestamp).total_seconds()
+        if dt <= 0:
+            raise ValueError(
+                f"new timestamp is not greater than the last timestamp; expecting: {timestamp} > {self._last_timestamp}"
+            )
+
+        dx = (keypoints - self._x_filtered) / dt
+
+        # Determine cutoff frequency based on movement speed
+        cutoff = self._min_cutoff + self._beta * jnp.linalg.norm(
+            dx, axis=-1, keepdims=True
+        )
+
+        # Apply low-pass filter to velocity
+        a_d = self._smoothing_factor(self._d_cutoff, dt)
+        self._dx_filtered = self._exponential_smoothing(
+            a_d,
+            dx,
+            (
+                jnp.zeros_like(keypoints)
+                if self._dx_filtered is None
+                else self._dx_filtered
+            ),
+        )
+
+        # Apply low-pass filter to position with adaptive cutoff
+        a_cutoff = self._smoothing_factor(jnp.asarray(cutoff), dt)
+        self._x_filtered = self._exponential_smoothing(
+            a_cutoff, keypoints, self._x_filtered
+        )
+
+        # Update timestamp
+        self._last_timestamp = timestamp
+
+    def get(self) -> TrackingPrediction:
+        """
+        Get the current state of the filter.
+
+        Returns:
+            TrackingPrediction with velocity and keypoints
+        """
+        return TrackingPrediction(
+            velocity=self._dx_filtered,
+            keypoints=self._x_filtered,
+        )
 
 
 @jaxtyped(typechecker=beartype)
 @dataclass(frozen=True)
-class Tracking:
-    id: int
+class TrackingState:
     """
-    The tracking id
+    immutable state of a tracking
     """
+
     keypoints: Float[Array, "J 3"]
     """
     The 3D keypoints of the tracking
@@ -41,50 +457,97 @@ class Tracking:
     The last active timestamp of the tracking
     """
 
-    historical_detections: PVector[Detection]
+    historical_detections_by_camera: PMap[CameraID, Detection]
     """
     Historical detections of the tracking.
 
     Used for 3D re-triangulation
     """
 
-    velocity: Optional[Float[Array, "3"]] = None
-    """
-    Could be `None`. Like when the 3D pose is initialized.
 
-    `velocity` should be updated when target association yields a new
-    3D pose.
-    """
+class Tracking:
+    id: TrackingID
+    state: TrackingState
+    velocity_filter: GenericVelocityFilter
+
+    def __init__(
+        self,
+        id: TrackingID,
+        state: TrackingState,
+        velocity_filter: Optional[GenericVelocityFilter] = None,
+    ):
+        self.id = id
+        self.state = state
+        self.velocity_filter = velocity_filter or DummyVelocityFilter(state.keypoints)
 
     def __repr__(self) -> str:
-        return f"Tracking({self.id}, {self.last_active_timestamp})"
+        return f"Tracking({self.id}, {self.state.last_active_timestamp})"
+
+    @overload
+    def predict(self, time: float) -> Float[Array, "J 3"]:
+        """
+        predict the keypoints at a given time
+
+        Args:
+            time: the time in seconds to predict the keypoints
+
+        Returns:
+            the predicted keypoints
+        """
+        ...  # pylint: disable=unnecessary-ellipsis
+
+    @overload
+    def predict(self, time: timedelta) -> Float[Array, "J 3"]:
+        """
+        predict the keypoints at a given time
+
+        Args:
+            time: the time delta to predict the keypoints
+        """
+        ...  # pylint: disable=unnecessary-ellipsis
+
+    @overload
+    def predict(self, time: datetime) -> Float[Array, "J 3"]:
+        """
+        predict the keypoints at a given time
+
+        Args:
+            time: the timestamp to predict the keypoints
+        """
+        ...  # pylint: disable=unnecessary-ellipsis
 
     def predict(
         self,
-        delta_t_s: float,
+        time: float | timedelta | datetime,
     ) -> Float[Array, "J 3"]:
-        """
-        Predict the 3D pose of a tracking based on its velocity.
-        JAX-friendly implementation that avoids Python control flow.
-
-        Args:
-            delta_t_s: Time delta in seconds
-
-        Returns:
-            Predicted 3D pose keypoints
-        """
-        # ------------------------------------------------------------------
-        # Step 1 – decide velocity on the Python side
-        # ------------------------------------------------------------------
-        if self.velocity is None:
-            velocity = jnp.zeros_like(self.keypoints)  # (J, 3)
+        if isinstance(time, timedelta):
+            timestamp = self.state.last_active_timestamp + time
+        elif isinstance(time, datetime):
+            timestamp = time
         else:
-            velocity = self.velocity  # (J, 3)
+            timestamp = self.state.last_active_timestamp + timedelta(seconds=time)
+        # pylint: disable-next=unsubscriptable-object
+        return self.velocity_filter.predict(timestamp)["keypoints"]
 
-        # ------------------------------------------------------------------
-        # Step 2 – pure JAX math
-        # ------------------------------------------------------------------
-        return self.keypoints + velocity * delta_t_s
+    def update(self, new_3d_pose: Float[Array, "J 3"], timestamp: datetime) -> None:
+        """
+        update the tracking with a new 3D pose
+
+        Note:
+            equivalent to call `velocity_filter.update(new_3d_pose, timestamp)`
+        """
+        self.velocity_filter.update(new_3d_pose, timestamp)
+
+    @property
+    def velocity(self) -> Float[Array, "J 3"]:
+        """
+        The velocity of the tracking for each keypoint
+        """
+        # pylint: disable-next=unsubscriptable-object
+        if (vel := self.velocity_filter.get()["velocity"]) is None:
+            return jnp.zeros_like(self.state.keypoints)
+        else:
+            return vel
 
 
 @jaxtyped(typechecker=beartype)
@@ -97,10 +560,10 @@ class AffinityResult:
     matrix: Float[Array, "T D"]
     trackings: Sequence[Tracking]
     detections: Sequence[Detection]
-    indices_T: Int[Array, "T"]  # pylint: disable=invalid-name
-    indices_D: Int[Array, "D"]  # pylint: disable=invalid-name
+    indices_T: Int[Array, "A"]  # pylint: disable=invalid-name
+    indices_D: Int[Array, "A"]  # pylint: disable=invalid-name
 
-    def tracking_detections(
+    def tracking_association(
         self,
     ) -> Generator[tuple[float, Tracking, Detection], None, None]:
         """

filter_object_by_box.py

@@ -0,0 +1,208 @@
+from narwhals import Boolean
+import numpy as np
+import cv2
+from typing import (
+    TypeAlias,
+    TypedDict,
+)
+from jaxtyping import Array, Num
+from shapely.geometry import Polygon
+from sympy import false, true
+
+NDArray: TypeAlias = np.ndarray
+
+
+# 盒子各个面的三维三角形集合
+box_triangles_list = [
+    ["4", "6", "7"],
+    ["4", "5", "6"],
+    ["2", "5", "6"],
+    ["1", "2", "5"],
+    ["1", "2", "3"],
+    ["0", "1", "3"],
+    ["0", "3", "7"],
+    ["0", "4", "7"],
+    ["2", "6", "7"],
+    ["2", "3", "7"],
+    ["0", "4", "5"],
+    ["0", "1", "5"],
+]
+
+
+class Camera_Params(TypedDict):
+    rvec: Num[NDArray, "3"]
+    tvec: Num[NDArray, "3"]
+    camera_matrix: Num[Array, "3 3"]
+    dist: Num[Array, "N"]
+    width: int
+    height: int
+
+
+class KeypointDataset(TypedDict):
+    frame_index: int
+    boxes: Num[NDArray, "N 4"]
+    kps: Num[NDArray, "N J 2"]
+    kps_scores: Num[NDArray, "N J"]
+
+
+# 三维坐标系根据相机内外参计算该镜头下的二维重投影坐标
+def reprojet_3d_to_2d(point_3d, camera_param):
+    point_2d, _ = cv2.projectPoints(
+        objectPoints=point_3d,
+        rvec=np.array(camera_param.params.Rt[:3, :3]),
+        tvec=np.array(camera_param.params.Rt[:3, 3]),
+        cameraMatrix=np.array(camera_param.params.K),
+        distCoeffs=np.array(camera_param.params.dist_coeffs),
+    )
+
+    point_2d = point_2d.reshape(-1).astype(int)
+    return point_2d
+
+
+# 计算盒子三维坐标系
+def calculaterCubeVersices(position, dimensions):
+    [cx, cy, cz] = position
+    [width, height, depth] = dimensions
+    halfWidth = width / 2
+    halfHeight = height / 2
+    halfDepth = depth / 2
+
+    return [
+        [cx - halfWidth, cy - halfHeight, cz - halfDepth],
+        [cx + halfWidth, cy - halfHeight, cz - halfDepth],
+        [cx + halfWidth, cy + halfHeight, cz - halfDepth],
+        [cx - halfWidth, cy + halfHeight, cz - halfDepth],
+        [cx - halfWidth, cy - halfHeight, cz + halfDepth],
+        [cx + halfWidth, cy - halfHeight, cz + halfDepth],
+        [cx + halfWidth, cy + halfHeight, cz + halfDepth],
+        [cx - halfWidth, cy + halfHeight, cz + halfDepth],
+    ]
+
+
+# 获得盒子三维坐标系
+def calculater_box_3d_points():
+    # 盒子原点位置，相对于六面体中心偏移
+    box_ori_potision = [0.205 + 0.2, 0.205 + 0.50, -0.205 - 0.45]
+    # 盒子边长，宽：1.5米，高：1.5米，深度：1.8米
+    box_geometry = [0.65, 1.8, 1]
+    filter_box_points_3d = calculaterCubeVersices(box_ori_potision, box_geometry)
+    filter_box_points_3d = {
+        str(index): element for index, element in enumerate(filter_box_points_3d)
+    }
+
+    return filter_box_points_3d
+
+
+# 计算盒子坐标系的二维重投影数据
+def calculater_box_2d_points(filter_box_points_3d, camera_param):
+    box_points_2d = dict()
+    for element_index, elment_point_3d in enumerate(filter_box_points_3d.values()):
+        box_points_2d[str(element_index)] = reprojet_3d_to_2d(
+            np.array(elment_point_3d), camera_param
+        ).tolist()
+
+    return box_points_2d
+
+
+# 盒子总的二维平面各三角形坐标点
+def calculater_box_common_scope(box_points_2d):
+    box_triangles_all_points = []
+    # 遍历三角形个数
+    for i in range(len(box_triangles_list)):
+        # 获取单个三角形二维平面坐标点
+        single_triangles = []
+        for element_key in box_triangles_list[i]:
+            single_triangles.append(box_points_2d[element_key])
+        box_triangles_all_points.append(single_triangles)
+
+    return box_triangles_all_points
+
+
+def calculate_triangle_union(triangles):
+    """
+    计算多个三角形的并集区域
+
+    参数:
+    triangles: 包含多个三角形的列表，每个三角形由三个点的坐标组成
+
+    返回:
+    union_area: 并集区域的面积
+    union_polygon: 表示并集区域的多边形对象
+    """
+    # 创建多边形对象列表
+    polygons = [Polygon(tri) for tri in triangles]
+
+    # 计算并集
+    union_polygon = polygons[0]
+    for polygon in polygons[1:]:
+        union_polygon = union_polygon.union(polygon)
+
+    # 计算并集面积
+    union_area = union_polygon.area
+
+    return union_area, union_polygon
+
+
+# 射线法，判断坐标点是否在box二维重投影的区域内
+def point_in_polygon(p, polygon):
+    x, y = p
+    n = len(polygon)
+    intersections = 0
+    on_boundary = False
+
+    for i in range(n):
+        xi, yi = polygon[i]
+        xj, yj = polygon[(i + 1) % n]  # 闭合多边形
+
+        # 检查点是否在顶点上
+        if (x == xi and y == yi) or (x == xj and y == yj):
+            on_boundary = True
+            break
+
+        # 检查点是否在线段上（非顶点情况）
+        if (min(xi, xj) <= x <= max(xi, xj)) and (min(yi, yj) <= y <= max(yi, yj)):
+            cross = (x - xi) * (yj - yi) - (y - yi) * (xj - xi)
+            if cross == 0:
+                on_boundary = True
+                break
+
+        # 计算射线与边的交点（非水平边）
+        if (yi > y) != (yj > y):
+            slope = (xj - xi) / (yj - yi) if (yj - yi) != 0 else float("inf")
+            x_intersect = xi + (y - yi) * slope
+            if x <= x_intersect:
+                intersections += 1
+
+    if on_boundary:
+        return false
+    return intersections % 2 == 1  # 奇数为内部，返回True
+
+
+# 获取并集区域坐标点
+def get_contours(union_polygon):
+    if union_polygon.geom_type == "Polygon":
+        # 单一多边形
+        x, y = union_polygon.exterior.xy
+
+        contours = [(list(x)[i], list(y)[i]) for i in range(len(x))]
+        contours = np.array(contours, np.int32)
+    return contours
+
+
+# 筛选落在盒子二维重投影区域内的关键点信息
+def filter_kps_in_contours(kps, contours) -> Boolean:
+
+    # 4 5 16 17
+    keypoint_index: list[list[int]] = [[4, 5], [16, 17]]
+    centers = []
+    for element_keypoint in keypoint_index:
+        x1, y1 = kps[element_keypoint[0]]
+        x2, y2 = kps[element_keypoint[1]]
+        centers.append([(x1 + x2) / 2, (y1 + y2) / 2])
+
+    if point_in_polygon(centers[0], contours) and point_in_polygon(
+        centers[1], contours
+    ):
+        return true
+    else:
+        return false

optical_detect_result/5606_demo.json

@@ -0,0 +1,282 @@
+[
+    {
+        "kps": [
+            419.0,
+            154.0
+        ],
+        "kps_scores": 1.0,
+        "index": 0
+    },
+    {
+        "kps": [
+            419.0521240234375,
+            154.07498168945312
+        ],
+        "kps_scores": 1.0,
+        "index": 1
+    },
+    {
+        "kps": [
+            418.5992736816406,
+            154.3507080078125
+        ],
+        "kps_scores": 1.0,
+        "index": 2
+    },
+    {
+        "kps": [
+            417.0777893066406,
+            154.17327880859375
+        ],
+        "kps_scores": 1.0,
+        "index": 3
+    },
+    {
+        "kps": [
+            416.8981628417969,
+            154.15330505371094
+        ],
+        "kps_scores": 1.0,
+        "index": 4
+    },
+    {
+        "kps": [
+            415.1317443847656,
+            153.68324279785156
+        ],
+        "kps_scores": 1.0,
+        "index": 5
+    },
+    {
+        "kps": [
+            413.2596130371094,
+            153.39761352539062
+        ],
+        "kps_scores": 1.0,
+        "index": 6
+    },
+    {
+        "kps": [
+            412.7089538574219,
+            153.3645782470703
+        ],
+        "kps_scores": 1.0,
+        "index": 7
+    },
+    {
+        "kps": [
+            409.3253173828125,
+            152.9347686767578
+        ],
+        "kps_scores": 1.0,
+        "index": 8
+    },
+    {
+        "kps": [
+            404.74853515625,
+            152.21153259277344
+        ],
+        "kps_scores": 1.0,
+        "index": 9
+    },
+    {
+        "kps": [
+            404.3977355957031,
+            152.19647216796875
+        ],
+        "kps_scores": 1.0,
+        "index": 10
+    },
+    {
+        "kps": [
+            396.53131103515625,
+            152.09912109375
+        ],
+        "kps_scores": 1.0,
+        "index": 11
+    },
+    {
+        "kps": [
+            393.76605224609375,
+            151.91282653808594
+        ],
+        "kps_scores": 1.0,
+        "index": 12
+    },
+    {
+        "kps": [
+            393.28106689453125,
+            151.76124572753906
+        ],
+        "kps_scores": 1.0,
+        "index": 13
+    },
+    {
+        "kps": [
+            383.2342834472656,
+            152.3790740966797
+        ],
+        "kps_scores": 1.0,
+        "index": 14
+    },
+    {
+        "kps": [
+            379.7545471191406,
+            152.79055786132812
+        ],
+        "kps_scores": 1.0,
+        "index": 15
+    },
+    {
+        "kps": [
+            379.8231506347656,
+            152.8155975341797
+        ],
+        "kps_scores": 1.0,
+        "index": 16
+    },
+    {
+        "kps": [
+            370.0028076171875,
+            155.16213989257812
+        ],
+        "kps_scores": 1.0,
+        "index": 17
+    },
+    {
+        "kps": [
+            366.5267639160156,
+            155.72059631347656
+        ],
+        "kps_scores": 1.0,
+        "index": 18
+    },
+    {
+        "kps": [
+            366.69610595703125,
+            156.3056182861328
+        ],
+        "kps_scores": 1.0,
+        "index": 19
+    },
+    {
+        "kps": [
+            359.8770751953125,
+            158.69798278808594
+        ],
+        "kps_scores": 1.0,
+        "index": 20
+    },
+    {
+        "kps": [
+            356.67681884765625,
+            160.0414581298828
+        ],
+        "kps_scores": 1.0,
+        "index": 21
+    },
+    {
+        "kps": [
+            348.1063232421875,
+            163.32858276367188
+        ],
+        "kps_scores": 1.0,
+        "index": 22
+    },
+    {
+        "kps": [
+            343.6862487792969,
+            165.0043182373047
+        ],
+        "kps_scores": 1.0,
+        "index": 23
+    },
+    {
+        "kps": [
+            339.2411804199219,
+            167.18580627441406
+        ],
+        "kps_scores": 1.0,
+        "index": 24
+    },
+    {
+        "kps": [
+            330.0,
+            170.0
+        ],
+        "kps_scores": 0.0,
+        "index": 25
+    },
+    {
+        "kps": [
+            322.0425720214844,
+            174.9293975830078
+        ],
+        "kps_scores": 1.0,
+        "index": 26
+    },
+    {
+        "kps": [
+            310.0,
+            176.0
+        ],
+        "kps_scores": 0.0,
+        "index": 27
+    },
+    {
+        "kps": [
+            305.0433349609375,
+            178.03123474121094
+        ],
+        "kps_scores": 1.0,
+        "index": 28
+    },
+    {
+        "kps": [
+            293.71295166015625,
+            183.8294219970703
+        ],
+        "kps_scores": 1.0,
+        "index": 29
+    },
+    {
+        "kps": [
+            291.28656005859375,
+            184.33445739746094
+        ],
+        "kps_scores": 1.0,
+        "index": 30
+    },
+    {
+        "kps": [
+            281.0,
+            190.0
+        ],
+        "kps_scores": 0.0,
+        "index": 31
+    },
+    {
+        "kps": [
+            272.0,
+            200.0
+        ],
+        "kps_scores": 0.0,
+        "index": 32
+    },
+    {
+        "kps": [
+            261.0457763671875,
+            211.67132568359375
+        ],
+        "kps_scores": 1.0,
+        "index": 33
+    },
+    {
+        "kps": [
+            239.03567504882812,
+            248.68519592285156
+        ],
+        "kps_scores": 1.0,
+        "index": 34
+    }
+]

optical_detect_result/5608_demo.json

@@ -0,0 +1,282 @@
+[
+    {
+        "kps": [
+            474.0,
+            215.00003051757812
+        ],
+        "kps_scores": 1.0,
+        "index": 0
+    },
+    {
+        "kps": [
+            474.0710754394531,
+            215.04542541503906
+        ],
+        "kps_scores": 1.0,
+        "index": 1
+    },
+    {
+        "kps": [
+            476.81365966796875,
+            215.0387420654297
+        ],
+        "kps_scores": 1.0,
+        "index": 2
+    },
+    {
+        "kps": [
+            479.3288269042969,
+            214.4371795654297
+        ],
+        "kps_scores": 1.0,
+        "index": 3
+    },
+    {
+        "kps": [
+            479.3817443847656,
+            214.49256896972656
+        ],
+        "kps_scores": 1.0,
+        "index": 4
+    },
+    {
+        "kps": [
+            483.0047302246094,
+            213.85231018066406
+        ],
+        "kps_scores": 1.0,
+        "index": 5
+    },
+    {
+        "kps": [
+            484.1208801269531,
+            213.64219665527344
+        ],
+        "kps_scores": 1.0,
+        "index": 6
+    },
+    {
+        "kps": [
+            484.140869140625,
+            213.63470458984375
+        ],
+        "kps_scores": 1.0,
+        "index": 7
+    },
+    {
+        "kps": [
+            487.458251953125,
+            213.45497131347656
+        ],
+        "kps_scores": 1.0,
+        "index": 8
+    },
+    {
+        "kps": [
+            488.8343505859375,
+            213.4651336669922
+        ],
+        "kps_scores": 1.0,
+        "index": 9
+    },
+    {
+        "kps": [
+            488.899658203125,
+            213.48526000976562
+        ],
+        "kps_scores": 1.0,
+        "index": 10
+    },
+    {
+        "kps": [
+            493.831787109375,
+            214.70533752441406
+        ],
+        "kps_scores": 1.0,
+        "index": 11
+    },
+    {
+        "kps": [
+            495.60980224609375,
+            215.26271057128906
+        ],
+        "kps_scores": 1.0,
+        "index": 12
+    },
+    {
+        "kps": [
+            495.5881042480469,
+            215.2436065673828
+        ],
+        "kps_scores": 1.0,
+        "index": 13
+    },
+    {
+        "kps": [
+            502.015380859375,
+            217.81201171875
+        ],
+        "kps_scores": 1.0,
+        "index": 14
+    },
+    {
+        "kps": [
+            504.2356262207031,
+            218.78392028808594
+        ],
+        "kps_scores": 1.0,
+        "index": 15
+    },
+    {
+        "kps": [
+            504.2625427246094,
+            218.81021118164062
+        ],
+        "kps_scores": 1.0,
+        "index": 16
+    },
+    {
+        "kps": [
+            511.97552490234375,
+            222.26150512695312
+        ],
+        "kps_scores": 1.0,
+        "index": 17
+    },
+    {
+        "kps": [
+            514.9180908203125,
+            224.3387908935547
+        ],
+        "kps_scores": 1.0,
+        "index": 18
+    },
+    {
+        "kps": [
+            514.7620239257812,
+            224.2892608642578
+        ],
+        "kps_scores": 1.0,
+        "index": 19
+    },
+    {
+        "kps": [
+            524.9593505859375,
+            230.30003356933594
+        ],
+        "kps_scores": 1.0,
+        "index": 20
+    },
+    {
+        "kps": [
+            528.3402709960938,
+            232.76568603515625
+        ],
+        "kps_scores": 1.0,
+        "index": 21
+    },
+    {
+        "kps": [
+            528.371826171875,
+            232.73399353027344
+        ],
+        "kps_scores": 1.0,
+        "index": 22
+    },
+    {
+        "kps": [
+            538.7906494140625,
+            240.9889678955078
+        ],
+        "kps_scores": 1.0,
+        "index": 23
+    },
+    {
+        "kps": [
+            538.7630004882812,
+            241.00299072265625
+        ],
+        "kps_scores": 1.0,
+        "index": 24
+    },
+    {
+        "kps": [
+            550.0248413085938,
+            248.24708557128906
+        ],
+        "kps_scores": 1.0,
+        "index": 25
+    },
+    {
+        "kps": [
+            554.3512573242188,
+            250.6501922607422
+        ],
+        "kps_scores": 1.0,
+        "index": 26
+    },
+    {
+        "kps": [
+            554.0921020507812,
+            250.47769165039062
+        ],
+        "kps_scores": 1.0,
+        "index": 27
+    },
+    {
+        "kps": [
+            567.93212890625,
+            266.1629943847656
+        ],
+        "kps_scores": 1.0,
+        "index": 28
+    },
+    {
+        "kps": [
+            571.8528442382812,
+            273.5104675292969
+        ],
+        "kps_scores": 1.0,
+        "index": 29
+    },
+    {
+        "kps": [
+            571.9888305664062,
+            273.5711669921875
+        ],
+        "kps_scores": 1.0,
+        "index": 30
+    },
+    {
+        "kps": [
+            586.6533203125,
+            309.09576416015625
+        ],
+        "kps_scores": 1.0,
+        "index": 31
+    },
+    {
+        "kps": [
+            591.8392944335938,
+            325.38385009765625
+        ],
+        "kps_scores": 1.0,
+        "index": 32
+    },
+    {
+        "kps": [
+            592.3212280273438,
+            325.2934265136719
+        ],
+        "kps_scores": 1.0,
+        "index": 33
+    },
+    {
+        "kps": [
+            603.3639526367188,
+            362.4980773925781
+        ],
+        "kps_scores": 1.0,
+        "index": 34
+    }
+]

optical_detect_result/5609_demo.json

@@ -0,0 +1,282 @@
+[
+    {
+        "kps": [
+            461.0,
+            164.0
+        ],
+        "kps_scores": 1.0,
+        "index": 0
+    },
+    {
+        "kps": [
+            460.9234619140625,
+            164.2275390625
+        ],
+        "kps_scores": 1.0,
+        "index": 1
+    },
+    {
+        "kps": [
+            460.93524169921875,
+            164.19480895996094
+        ],
+        "kps_scores": 1.0,
+        "index": 2
+    },
+    {
+        "kps": [
+            460.4592590332031,
+            164.14320373535156
+        ],
+        "kps_scores": 1.0,
+        "index": 3
+    },
+    {
+        "kps": [
+            459.9245910644531,
+            164.054931640625
+        ],
+        "kps_scores": 1.0,
+        "index": 4
+    },
+    {
+        "kps": [
+            459.8656921386719,
+            164.08154296875
+        ],
+        "kps_scores": 1.0,
+        "index": 5
+    },
+    {
+        "kps": [
+            456.9087219238281,
+            163.1707305908203
+        ],
+        "kps_scores": 1.0,
+        "index": 6
+    },
+    {
+        "kps": [
+            455.7566223144531,
+            162.69784545898438
+        ],
+        "kps_scores": 1.0,
+        "index": 7
+    },
+    {
+        "kps": [
+            455.740478515625,
+            162.74818420410156
+        ],
+        "kps_scores": 1.0,
+        "index": 8
+    },
+    {
+        "kps": [
+            449.8667907714844,
+            161.95462036132812
+        ],
+        "kps_scores": 1.0,
+        "index": 9
+    },
+    {
+        "kps": [
+            447.55975341796875,
+            162.12559509277344
+        ],
+        "kps_scores": 1.0,
+        "index": 10
+    },
+    {
+        "kps": [
+            447.5325012207031,
+            162.12460327148438
+        ],
+        "kps_scores": 1.0,
+        "index": 11
+    },
+    {
+        "kps": [
+            439.9998474121094,
+            162.59873962402344
+        ],
+        "kps_scores": 1.0,
+        "index": 12
+    },
+    {
+        "kps": [
+            437.3090515136719,
+            162.88577270507812
+        ],
+        "kps_scores": 1.0,
+        "index": 13
+    },
+    {
+        "kps": [
+            437.2088623046875,
+            162.84994506835938
+        ],
+        "kps_scores": 1.0,
+        "index": 14
+    },
+    {
+        "kps": [
+            429.199951171875,
+            164.5860595703125
+        ],
+        "kps_scores": 1.0,
+        "index": 15
+    },
+    {
+        "kps": [
+            429.32745361328125,
+            164.66001892089844
+        ],
+        "kps_scores": 1.0,
+        "index": 16
+    },
+    {
+        "kps": [
+            424.8293762207031,
+            166.40106201171875
+        ],
+        "kps_scores": 1.0,
+        "index": 17
+    },
+    {
+        "kps": [
+            419.6496887207031,
+            168.80294799804688
+        ],
+        "kps_scores": 1.0,
+        "index": 18
+    },
+    {
+        "kps": [
+            419.6795349121094,
+            168.93418884277344
+        ],
+        "kps_scores": 1.0,
+        "index": 19
+    },
+    {
+        "kps": [
+            414.8919677734375,
+            172.65428161621094
+        ],
+        "kps_scores": 1.0,
+        "index": 20
+    },
+    {
+        "kps": [
+            410.0992431640625,
+            175.77218627929688
+        ],
+        "kps_scores": 1.0,
+        "index": 21
+    },
+    {
+        "kps": [
+            410.0442810058594,
+            175.911376953125
+        ],
+        "kps_scores": 1.0,
+        "index": 22
+    },
+    {
+        "kps": [
+            400.20159912109375,
+            184.33380126953125
+        ],
+        "kps_scores": 1.0,
+        "index": 23
+    },
+    {
+        "kps": [
+            396.4606628417969,
+            186.7172088623047
+        ],
+        "kps_scores": 1.0,
+        "index": 24
+    },
+    {
+        "kps": [
+            396.3185119628906,
+            186.76808166503906
+        ],
+        "kps_scores": 1.0,
+        "index": 25
+    },
+    {
+        "kps": [
+            382.623291015625,
+            192.941650390625
+        ],
+        "kps_scores": 1.0,
+        "index": 26
+    },
+    {
+        "kps": [
+            376.8236999511719,
+            195.2269744873047
+        ],
+        "kps_scores": 1.0,
+        "index": 27
+    },
+    {
+        "kps": [
+            376.66937255859375,
+            195.1109161376953
+        ],
+        "kps_scores": 1.0,
+        "index": 28
+    },
+    {
+        "kps": [
+            362.7231750488281,
+            209.30923461914062
+        ],
+        "kps_scores": 1.0,
+        "index": 29
+    },
+    {
+        "kps": [
+            355.9901123046875,
+            216.26303100585938
+        ],
+        "kps_scores": 1.0,
+        "index": 30
+    },
+    {
+        "kps": [
+            356.3956298828125,
+            216.3310546875
+        ],
+        "kps_scores": 1.0,
+        "index": 31
+    },
+    {
+        "kps": [
+            343.6780090332031,
+            235.2663116455078
+        ],
+        "kps_scores": 1.0,
+        "index": 32
+    },
+    {
+        "kps": [
+            332.50238037109375,
+            261.8990783691406
+        ],
+        "kps_scores": 1.0,
+        "index": 33
+    },
+    {
+        "kps": [
+            332.8721923828125,
+            261.7060546875
+        ],
+        "kps_scores": 1.0,
+        "index": 34
+    }
+]

playground.py

@@ -31,13 +31,13 @@ from typing import (
     TypeVar,
     cast,
     overload,
+    Iterable,
 )
 
 import awkward as ak
 import jax
 import jax.numpy as jnp
 import numpy as np
-import orjson
 from beartype import beartype
 from beartype.typing import Mapping, Sequence
 from cv2 import undistortPoints
@@ -46,9 +46,10 @@ from jaxtyping import Array, Float, Num, jaxtyped
 from matplotlib import pyplot as plt
 from numpy.typing import ArrayLike
 from optax.assignment import hungarian_algorithm as linear_sum_assignment
-from pyrsistent import v, pvector
+from pyrsistent import pvector, v, m, pmap, PMap, freeze, thaw
 from scipy.spatial.transform import Rotation as R
 from typing_extensions import deprecated
+from collections import defaultdict
 
 from app.camera import (
     Camera,
@@ -59,15 +60,21 @@ from app.camera import (
     classify_by_camera,
 )
 from app.solver._old import GLPKSolver
-from app.tracking import AffinityResult, Tracking
+from app.tracking import (
+    TrackingID,
+    AffinityResult,
+    LastDifferenceVelocityFilter,
+    Tracking,
+    TrackingState,
+)
 from app.visualize.whole_body import visualize_whole_body
 
 NDArray: TypeAlias = np.ndarray
 
 # %%
 DATASET_PATH = Path("samples") / "04_02"
-AK_CAMERA_DATASET: ak.Array = ak.from_parquet(DATASET_PATH / "camera_params.parquet")
-DELTA_T_MIN = timedelta(milliseconds=10)
+AK_CAMERA_DATASET: ak.Array = ak.from_parquet(DATASET_PATH / "camera_params.parquet")  # type: ignore
+DELTA_T_MIN = timedelta(milliseconds=1)
 display(AK_CAMERA_DATASET)
 
 
@@ -504,6 +511,142 @@ def triangulate_points_from_multiple_views_linear(
     return vmap_triangulate(proj_matrices, points, conf)
 
 
+# %%
+@jaxtyped(typechecker=beartype)
+def triangulate_one_point_from_multiple_views_linear_time_weighted(
+    proj_matrices: Float[Array, "N 3 4"],
+    points: Num[Array, "N 2"],
+    delta_t: Num[Array, "N"],
+    lambda_t: float = 10.0,
+    confidences: Optional[Float[Array, "N"]] = None,
+) -> Float[Array, "3"]:
+    """
+    Triangulate one point from multiple views with time-weighted linear least squares.
+
+    Implements the incremental reconstruction method from "Cross-View Tracking for Multi-Human 3D Pose"
+    with weighting formula: w_i = exp(-λ_t(t-t_i)) / ||c^i^T||_2
+
+    Args:
+        proj_matrices: Shape (N, 3, 4) projection matrices sequence
+        points: Shape (N, 2) point coordinates sequence
+        delta_t: Time differences between current time and each observation (in seconds)
+        lambda_t: Time penalty rate (higher values decrease influence of older observations)
+        confidences: Shape (N,) confidence values in range [0.0, 1.0]
+
+    Returns:
+        point_3d: Shape (3,) triangulated 3D point
+    """
+    assert len(proj_matrices) == len(points)
+    assert len(delta_t) == len(points)
+
+    N = len(proj_matrices)
+
+    # Prepare confidence weights
+    confi: Float[Array, "N"]
+    if confidences is None:
+        confi = jnp.ones(N, dtype=np.float32)
+    else:
+        confi = jnp.sqrt(jnp.clip(confidences, 0, 1))
+
+    A = jnp.zeros((N * 2, 4), dtype=np.float32)
+
+    # First build the coefficient matrix without weights
+    for i in range(N):
+        x, y = points[i]
+        A = A.at[2 * i].set(proj_matrices[i, 2] * x - proj_matrices[i, 0])
+        A = A.at[2 * i + 1].set(proj_matrices[i, 2] * y - proj_matrices[i, 1])
+
+    # Then apply the time-based and confidence weights
+    for i in range(N):
+        # Calculate time-decay weight: e^(-λ_t * Δt)
+        time_weight = jnp.exp(-lambda_t * delta_t[i])
+
+        # Calculate normalization factor: ||c^i^T||_2
+        row_norm_1 = jnp.linalg.norm(A[2 * i])
+        row_norm_2 = jnp.linalg.norm(A[2 * i + 1])
+
+        # Apply combined weight: time_weight / row_norm * confidence
+        w1 = (time_weight / row_norm_1) * confi[i]
+        w2 = (time_weight / row_norm_2) * confi[i]
+
+        A = A.at[2 * i].mul(w1)
+        A = A.at[2 * i + 1].mul(w2)
+
+    # Solve using SVD
+    _, _, vh = jnp.linalg.svd(A, full_matrices=False)
+    point_3d_homo = vh[-1]  # shape (4,)
+
+    # Ensure homogeneous coordinate is positive
+    point_3d_homo = jnp.where(
+        point_3d_homo[3] < 0,
+        -point_3d_homo,
+        point_3d_homo,
+    )
+
+    # Convert from homogeneous to Euclidean coordinates
+    point_3d = point_3d_homo[:3] / point_3d_homo[3]
+    return point_3d
+
+
+@jaxtyped(typechecker=beartype)
+def triangulate_points_from_multiple_views_linear_time_weighted(
+    proj_matrices: Float[Array, "N 3 4"],
+    points: Num[Array, "N P 2"],
+    delta_t: Num[Array, "N"],
+    lambda_t: float = 10.0,
+    confidences: Optional[Float[Array, "N P"]] = None,
+) -> Float[Array, "P 3"]:
+    """
+    Vectorized version that triangulates P points from N camera views with time-weighting.
+
+    This function uses JAX's vmap to efficiently triangulate multiple points in parallel.
+
+    Args:
+        proj_matrices: Shape (N, 3, 4) projection matrices for N cameras
+        points: Shape (N, P, 2) 2D points for P keypoints across N cameras
+        delta_t: Shape (N,) time differences between current time and each camera's timestamp (seconds)
+        lambda_t: Time penalty rate (higher values decrease influence of older observations)
+        confidences: Shape (N, P) confidence values for each point in each camera
+
+    Returns:
+        points_3d: Shape (P, 3) triangulated 3D points
+    """
+    N, P, _ = points.shape
+    assert (
+        proj_matrices.shape[0] == N
+    ), "Number of projection matrices must match number of cameras"
+    assert delta_t.shape[0] == N, "Number of time deltas must match number of cameras"
+
+    if confidences is None:
+        # Create uniform confidences if none provided
+        conf = jnp.ones((N, P), dtype=jnp.float32)
+    else:
+        conf = confidences
+
+    # Define the vmapped version of the single-point function
+    # We map over the second dimension (P points) of the input arrays
+    vmap_triangulate = jax.vmap(
+        triangulate_one_point_from_multiple_views_linear_time_weighted,
+        in_axes=(
+            None,
+            1,
+            None,
+            None,
+            1,
+        ),  # proj_matrices and delta_t static, map over points
+        out_axes=0,  # Output has first dimension corresponding to points
+    )
+
+    # For each point p, extract the 2D coordinates from all cameras and triangulate
+    return vmap_triangulate(
+        proj_matrices,  # (N, 3, 4) - static across points
+        points,  # (N, P, 2) - map over dim 1 (P)
+        delta_t,  # (N,) - static across points
+        lambda_t,  # scalar - static
+        conf,  # (N, P) - map over dim 1 (P)
+    )
+
+
 # %%
 
 
@@ -524,6 +667,21 @@ def triangle_from_cluster(
 
 
 # %%
+def group_by_cluster_by_camera(
+    cluster: Sequence[Detection],
+) -> PMap[CameraID, Detection]:
+    """
+    group the detections by camera, and preserve the latest detection for each camera
+    """
+    r: dict[CameraID, Detection] = {}
+    for el in cluster:
+        if el.camera.id in r:
+            eld = r[el.camera.id]
+            preserved = max([eld, el], key=lambda x: x.timestamp)
+            r[el.camera.id] = preserved
+    return pmap(r)
+
+
 class GlobalTrackingState:
     _last_id: int
     _trackings: dict[int, Tracking]
@@ -542,13 +700,21 @@ class GlobalTrackingState:
         return shallow_copy(self._trackings)
 
     def add_tracking(self, cluster: Sequence[Detection]) -> Tracking:
+        if len(cluster) < 2:
+            raise ValueError(
+                "cluster must contain at least 2 detections to form a tracking"
+            )
         kps_3d, latest_timestamp = triangle_from_cluster(cluster)
         next_id = self._last_id + 1
-        tracking = Tracking(
-            id=next_id,
+        tracking_state = TrackingState(
             keypoints=kps_3d,
             last_active_timestamp=latest_timestamp,
-            historical_detections=v(*cluster),
+            historical_detections_by_camera=group_by_cluster_by_camera(cluster),
+        )
+        tracking = Tracking(
+            id=next_id,
+            state=tracking_state,
+            velocity_filter=LastDifferenceVelocityFilter(kps_3d, latest_timestamp),
         )
         self._trackings[next_id] = tracking
         self._last_id = next_id
@@ -671,11 +837,7 @@ def perpendicular_distance_camera_2d_points_to_tracking_raycasting(
         Array of perpendicular distances for each keypoint
     """
     camera = detection.camera
-    # Use the delta_t supplied by the caller, but clamp to DELTA_T_MIN to
-    # avoid division-by-zero / exploding affinities.
-    delta_t = max(delta_t, DELTA_T_MIN)
-    delta_t_s = delta_t.total_seconds()
-    predicted_pose = tracking.predict(delta_t_s)
+    predicted_pose = tracking.predict(delta_t)
 
     # Back-project the 2D points to 3D space
     # intersection with z=0 plane
@@ -755,12 +917,12 @@ def calculate_tracking_detection_affinity(
         Combined affinity score
     """
     camera = detection.camera
-    delta_t_raw = detection.timestamp - tracking.last_active_timestamp
+    delta_t_raw = detection.timestamp - tracking.state.last_active_timestamp
     # Clamp delta_t to avoid division-by-zero / exploding affinity.
     delta_t = max(delta_t_raw, DELTA_T_MIN)
 
     # Calculate 2D affinity
-    tracking_2d_projection = camera.project(tracking.keypoints)
+    tracking_2d_projection = camera.project(tracking.state.keypoints)
     w, h = camera.params.image_size
     distance_2d = calculate_distance_2d(
         tracking_2d_projection,
@@ -840,7 +1002,7 @@ def calculate_camera_affinity_matrix_jax(
 
     # === Tracking-side tensors ===
     kps3d_trk: Float[Array, "T J 3"] = jnp.stack(
-        [trk.keypoints for trk in trackings]
+        [trk.state.keypoints for trk in trackings]
     )  # (T, J, 3)
     J = kps3d_trk.shape[1]
     # === Detection-side tensors ===
@@ -857,12 +1019,12 @@ def calculate_camera_affinity_matrix_jax(
     # --- timestamps ----------
     t0 = min(
         chain(
-            (trk.last_active_timestamp for trk in trackings),
+            (trk.state.last_active_timestamp for trk in trackings),
             (det.timestamp for det in camera_detections),
         )
     ).timestamp()  # common origin (float)
     ts_trk = jnp.array(
-        [trk.last_active_timestamp.timestamp() - t0 for trk in trackings],
+        [trk.state.last_active_timestamp.timestamp() - t0 for trk in trackings],
         dtype=jnp.float32,  # now small, ms-scale fits in fp32
     )
     ts_det = jnp.array(
@@ -1033,8 +1195,82 @@ display(affinities)
 
 
 # %%
-def update_tracking(tracking: Tracking, detection: Detection):
-    delta_t_ = detection.timestamp - tracking.last_active_timestamp
-    delta_t = max(delta_t_, DELTA_T_MIN)
+def affinity_result_by_tracking(
+    results: Iterable[AffinityResult],
+    min_affinity: float = 0.0,
+) -> dict[TrackingID, list[Detection]]:
+    """
+    Group affinity results by target ID.
 
-    return tracking
+    Args:
+        results: the affinity results to group
+        min_affinity: the minimum affinity to consider
+    Returns:
+        a dictionary mapping tracking IDs to a list of detections
+    """
+    res: dict[TrackingID, list[Detection]] = defaultdict(list)
+    for affinity_result in results:
+        for affinity, t, d in affinity_result.tracking_association():
+            if affinity < min_affinity:
+                continue
+            res[t.id].append(d)
+    return res
+
+
+def update_tracking(
+    tracking: Tracking,
+    detections: Sequence[Detection],
+    max_delta_t: timedelta = timedelta(milliseconds=100),
+    lambda_t: float = 10.0,
+) -> None:
+    """
+    update the tracking with a new set of detections
+
+    Args:
+        tracking: the tracking to update
+        detections: the detections to update the tracking with
+        max_delta_t: the maximum time difference between the last active timestamp and the latest detection
+        lambda_t: the lambda value for the time difference
+
+    Note:
+        the function would mutate the tracking object
+    """
+    last_active_timestamp = tracking.state.last_active_timestamp
+    latest_timestamp = max(d.timestamp for d in detections)
+    d = thaw(tracking.state.historical_detections_by_camera)
+    for detection in detections:
+        d[detection.camera.id] = detection
+    for camera_id, detection in d.items():
+        if detection.timestamp - latest_timestamp > max_delta_t:
+            del d[camera_id]
+    new_detections = freeze(d)
+    new_detections_list = list(new_detections.values())
+    project_matrices = jnp.stack(
+        [detection.camera.params.projection_matrix for detection in new_detections_list]
+    )
+    delta_t = jnp.array(
+        [
+            detection.timestamp.timestamp() - last_active_timestamp.timestamp()
+            for detection in new_detections_list
+        ]
+    )
+    kps = jnp.stack([detection.keypoints for detection in new_detections_list])
+    conf = jnp.stack([detection.confidences for detection in new_detections_list])
+    kps_3d = triangulate_points_from_multiple_views_linear_time_weighted(
+        project_matrices, kps, delta_t, lambda_t, conf
+    )
+    new_state = TrackingState(
+        keypoints=kps_3d,
+        last_active_timestamp=latest_timestamp,
+        historical_detections_by_camera=new_detections,
+    )
+    tracking.update(kps_3d, latest_timestamp)
+    tracking.state = new_state
+
+
+# %%
+affinity_results_by_tracking = affinity_result_by_tracking(affinities.values())
+for tracking_id, detections in affinity_results_by_tracking.items():
+    update_tracking(global_tracking_state.trackings[tracking_id], detections)
+
+# %%

pyproject.toml

@@ -14,6 +14,7 @@ dependencies = [
     "jaxtyping>=0.2.38",
     "jupytext>=1.17.0",
     "matplotlib>=3.10.1",
+    "more-itertools>=10.7.0",
     "opencv-python-headless>=4.11.0.86",
     "optax>=0.2.4",
     "orjson>=3.10.15",
@@ -23,6 +24,7 @@ dependencies = [
     "pyrsistent>=0.20.0",
     "pytest>=8.3.5",
     "scipy>=1.15.2",
+    "shapely>=2.1.1",
     "torch>=2.6.0",
     "torchvision>=0.21.0",
     "typeguard>=4.4.2",

smooth_3d_kps.ipynb

@@ -0,0 +1,122 @@
+{
+ "cells": [
+  {
+   "cell_type": "code",
+   "execution_count": 16,
+   "id": "0d48b7eb",
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "import json\n",
+    "from pathlib import Path\n",
+    "import numpy as np"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 17,
+   "id": "dfd27584",
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "KPS_PATH = Path(\"samples/WeiHua_03.json\")\n",
+    "with open(KPS_PATH, \"r\") as file:\n",
+    "    data = json.load(file)"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 18,
+   "id": "360f9c50",
+   "metadata": {},
+   "outputs": [
+    {
+     "data": {
+      "text/plain": [
+       "'index:1, shape: (33, 133, 3)'"
+      ]
+     },
+     "metadata": {},
+     "output_type": "display_data"
+    },
+    {
+     "data": {
+      "text/plain": [
+       "'index:2, shape: (662, 133, 3)'"
+      ]
+     },
+     "metadata": {},
+     "output_type": "display_data"
+    }
+   ],
+   "source": [
+    "for item_object_index in data.keys():\n",
+    "    item_object = np.array(data[item_object_index])\n",
+    "    display(f'index:{item_object_index}, shape: {item_object.shape}')"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "# 对data['2']的662帧3d关键点数据进行滑动窗口平滑处理\n",
+    "object_points = np.array(data['2'])  # shape: (662, 133, 3)\n",
+    "window_size = 5\n",
+    "kernel = np.ones(window_size) / window_size\n",
+    "# 对每个关键点的每个坐标轴分别做滑动平均\n",
+    "smoothed_points = np.zeros_like(object_points)\n",
+    "# 遍历133个关节\n",
+    "for kp_idx in range(object_points.shape[1]):\n",
+    "    # 遍历每个关节的空间三维坐标点\n",
+    "    for axis in range(3):\n",
+    "        # 对第i帧的滑动平滑方式 smoothed[i] = (point[i-2] + point[i-1] + point[i] + point[i+1] + point[i+2]) / 5\n",
+    "        smoothed_points[:, kp_idx, axis] = np.convolve(object_points[:, kp_idx, axis], kernel, mode='same')"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 20,
+   "id": "24c6c0c9",
+   "metadata": {},
+   "outputs": [
+    {
+     "data": {
+      "text/plain": [
+       "'smoothed_points shape: (662, 133, 3)'"
+      ]
+     },
+     "metadata": {},
+     "output_type": "display_data"
+    }
+   ],
+   "source": [
+    "display(f'smoothed_points shape: {smoothed_points.shape}')\n",
+    "with open(\"samples/smoothed_3d_kps.json\", \"w\") as file:\n",
+    "    json.dump({'1':smoothed_points.tolist()}, file)"
+   ]
+  }
+ ],
+ "metadata": {
+  "kernelspec": {
+   "display_name": "cvth3pe",
+   "language": "python",
+   "name": "python3"
+  },
+  "language_info": {
+   "codemirror_mode": {
+    "name": "ipython",
+    "version": 3
+   },
+   "file_extension": ".py",
+   "mimetype": "text/x-python",
+   "name": "python",
+   "nbconvert_exporter": "python",
+   "pygments_lexer": "ipython3",
+   "version": "3.12.9"
+  }
+ },
+ "nbformat": 4,
+ "nbformat_minor": 5
+}

temp_test_camera_ex_params.ipynb

@@ -0,0 +1,193 @@
+{
+ "cells": [
+  {
+   "cell_type": "code",
+   "execution_count": 8,
+   "id": "11cc2345",
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "import awkward as ak\n",
+    "import numpy as np\n",
+    "from pathlib import Path"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 9,
+   "id": "84348d97",
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "CAMERA_INDEX ={\n",
+    "    2:\"5602\",\n",
+    "    4:\"5604\",\n",
+    "}\n",
+    "index = 4\n",
+    "CAMERA_PATH = Path(\"/home/admin/Documents/ActualTest_QuanCheng/camera_ex_params_1_2025_4_20/camera_params\")\n",
+    "camera_data = ak.from_parquet(CAMERA_PATH / CAMERA_INDEX[index]/ \"extrinsic.parquet\")"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 10,
+   "id": "1d771740",
+   "metadata": {},
+   "outputs": [
+    {
+     "data": {
+      "text/html": [
+       "<pre>[{rvec: [[-2.26], [0.0669], [-2.15]], tvec: [[0.166], ...]},\n",
+       " {rvec: [[2.07], [0.144], [2.21]], tvec: [[0.143], ...]},\n",
+       " {rvec: [[2.09], [0.0872], [2.25]], tvec: [[0.141], ...]},\n",
+       " {rvec: [[2.16], [0.172], [2.09]], tvec: [[0.162], ...]},\n",
+       " {rvec: [[2.15], [0.18], [2.09]], tvec: [[0.162], ...]},\n",
+       " {rvec: [[-2.22], [0.117], [-2.14]], tvec: [[0.162], ...]},\n",
+       " {rvec: [[2.18], [0.176], [2.08]], tvec: [[0.166], ...]},\n",
+       " {rvec: [[2.18], [0.176], [2.08]], tvec: [[0.166], ...]},\n",
+       " {rvec: [[-2.26], [0.116], [-2.1]], tvec: [[0.17], ...]},\n",
+       " {rvec: [[-2.26], [0.124], [-2.09]], tvec: [[0.171], ...]},\n",
+       " ...,\n",
+       " {rvec: [[-2.2], [0.0998], [-2.17]], tvec: [[0.158], ...]},\n",
+       " {rvec: [[-2.2], [0.0998], [-2.17]], tvec: [[0.158], ...]},\n",
+       " {rvec: [[2.12], [0.151], [2.16]], tvec: [[0.152], ...]},\n",
+       " {rvec: [[-2.3], [0.0733], [-2.1]], tvec: [[0.175], ...]},\n",
+       " {rvec: [[2.1], [0.16], [2.17]], tvec: [[0.149], ...]},\n",
+       " {rvec: [[2.1], [0.191], [2.13]], tvec: [[0.153], ...]},\n",
+       " {rvec: [[2.11], [0.196], [2.12]], tvec: [[0.154], ...]},\n",
+       " {rvec: [[2.19], [0.171], [2.08]], tvec: [[0.166], ...]},\n",
+       " {rvec: [[2.24], [0.0604], [2.12]], tvec: [[0.166], ...]}]\n",
+       "---------------------------------------------------------------------------\n",
+       "backend: cpu\n",
+       "nbytes: 10.1 kB\n",
+       "type: 90 * {\n",
+       "    rvec: var * var * float64,\n",
+       "    tvec: var * var * float64\n",
+       "}</pre>"
+      ],
+      "text/plain": [
+       "<Array [{rvec: [...], tvec: [...]}, ..., {...}] type='90 * {rvec: var * var...'>"
+      ]
+     },
+     "metadata": {},
+     "output_type": "display_data"
+    }
+   ],
+   "source": [
+    "display(camera_data)"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 11,
+   "id": "59fde11b",
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "data = []\n",
+    "for element in camera_data:\n",
+    "    rvec = element[\"rvec\"]\n",
+    "    if rvec[0]<0:\n",
+    "        data.append({\"rvec\": rvec, \"tvec\": element[\"tvec\"]})"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 12,
+   "id": "4792cbc4",
+   "metadata": {},
+   "outputs": [
+    {
+     "data": {
+      "text/plain": [
+       "<pyarrow._parquet.FileMetaData object at 0x7799cbf62d40>\n",
+       "  created_by: parquet-cpp-arrow version 19.0.1\n",
+       "  num_columns: 2\n",
+       "  num_rows: 30\n",
+       "  num_row_groups: 1\n",
+       "  format_version: 2.6\n",
+       "  serialized_size: 0"
+      ]
+     },
+     "execution_count": 12,
+     "metadata": {},
+     "output_type": "execute_result"
+    }
+   ],
+   "source": [
+    "ak.to_parquet(ak.from_iter(data),\"/home/admin/Documents/ActualTest_QuanCheng/camera_ex_params_1_2025_4_20/camera_params/5604/re_extrinsic.parquet\")"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 13,
+   "id": "8225ee33",
+   "metadata": {},
+   "outputs": [
+    {
+     "data": {
+      "text/html": [
+       "<pre>[{rvec: [[-2.26], [0.0669], [-2.15]], tvec: [[0.166], ...]},\n",
+       " {rvec: [[-2.22], [0.117], [-2.14]], tvec: [[0.162], ...]},\n",
+       " {rvec: [[-2.26], [0.116], [-2.1]], tvec: [[0.17], ...]},\n",
+       " {rvec: [[-2.26], [0.124], [-2.09]], tvec: [[0.171], ...]},\n",
+       " {rvec: [[-2.24], [0.133], [-2.11]], tvec: [[0.167], ...]},\n",
+       " {rvec: [[-2.22], [0.0556], [-2.2]], tvec: [[0.158], ...]},\n",
+       " {rvec: [[-2.27], [0.119], [-2.09]], tvec: [[0.172], ...]},\n",
+       " {rvec: [[-2.34], [0.0663], [-2.06]], tvec: [[0.181], ...]},\n",
+       " {rvec: [[-2.21], [0.117], [-2.15]], tvec: [[0.161], ...]},\n",
+       " {rvec: [[-2.33], [0.0731], [-2.08]], tvec: [[0.179], ...]},\n",
+       " ...,\n",
+       " {rvec: [[-2.23], [0.106], [-2.13]], tvec: [[0.166], ...]},\n",
+       " {rvec: [[-2.21], [0.054], [-2.2]], tvec: [[0.157], ...]},\n",
+       " {rvec: [[-2.19], [0.0169], [-2.25]], tvec: [[0.151], ...]},\n",
+       " {rvec: [[-2.2], [0.0719], [-2.19]], tvec: [[0.157], ...]},\n",
+       " {rvec: [[-2.22], [0.0726], [-2.18]], tvec: [[0.161], ...]},\n",
+       " {rvec: [[-2.2], [0.0742], [-2.19]], tvec: [[0.158], ...]},\n",
+       " {rvec: [[-2.2], [0.0998], [-2.17]], tvec: [[0.158], ...]},\n",
+       " {rvec: [[-2.2], [0.0998], [-2.17]], tvec: [[0.158], ...]},\n",
+       " {rvec: [[-2.3], [0.0733], [-2.1]], tvec: [[0.175], ...]}]\n",
+       "---------------------------------------------------------------------------\n",
+       "backend: cpu\n",
+       "nbytes: 3.4 kB\n",
+       "type: 30 * {\n",
+       "    rvec: var * var * float64,\n",
+       "    tvec: var * var * float64\n",
+       "}</pre>"
+      ],
+      "text/plain": [
+       "<Array [{rvec: [...], tvec: [...]}, ..., {...}] type='30 * {rvec: var * var...'>"
+      ]
+     },
+     "metadata": {},
+     "output_type": "display_data"
+    }
+   ],
+   "source": [
+    "temp_data = ak.from_parquet(\"/home/admin/Documents/ActualTest_QuanCheng/camera_ex_params_1_2025_4_20/camera_params/5604/re_extrinsic.parquet\")\n",
+    "display(temp_data)"
+   ]
+  }
+ ],
+ "metadata": {
+  "kernelspec": {
+   "display_name": ".venv",
+   "language": "python",
+   "name": "python3"
+  },
+  "language_info": {
+   "codemirror_mode": {
+    "name": "ipython",
+    "version": 3
+   },
+   "file_extension": ".py",
+   "mimetype": "text/x-python",
+   "name": "python",
+   "nbconvert_exporter": "python",
+   "pygments_lexer": "ipython3",
+   "version": "3.12.9"
+  }
+ },
+ "nbformat": 4,
+ "nbformat_minor": 5
+}