HQU-gxy/CVTH3PE
7
0
Fork 2
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

add annotations

Browse Source
This commit is contained in:
crosstyan
2025-07-09 10:31:22 +08:00
parent b4ac324d8f
commit 0d16925741
1 changed files with 32 additions and 19 deletions
Download Patch File Download Diff File Expand all files Collapse all files

51
single_people_detect_track.py
View File

@ -976,44 +976,54 @@ tracking_initialized = False
lost_frame_count = 0
lost_frame_threshold = 12 # 0.5秒，假设20fps
# ===================== 主循环：逐帧处理检测与跟踪 =====================
while True:
# 重新梳理跟踪逻辑，保证唯一目标、唯一初始化、鲁棒丢失判定
try:
# 获取下一个时间步的所有相机检测结果
detections = next(sync_gen)
# 过滤低置信度的检测，提升后续三角化和跟踪的准确性
detections = filter_keypoints_by_scores(detections, threshold=0.5)
except StopIteration:
# 检测数据读取完毕，退出主循环
break
# 1. 检查当前是否有已初始化的跟踪目标
# 跟踪目标按id排序，便于后续一致性处理
trackings: list[Tracking] = sorted(
global_tracking_state.trackings.values(), key=lambda x: x.id
)
# ========== 跟踪目标初始化流程 ==========
if not tracking_initialized:
# 只初始化一次
camera_port_this = filter_camera_port(detections)
# 只初始化一次跟踪目标，防止多次重复初始化
camera_port_this = filter_camera_port(detections) # 获取当前帧检测到的相机端口
# 如果检测到的相机数量小于总机位数-1，则认为初始化条件不满足，跳过本帧
if len(camera_port_this) < len(camera_port) - 1:
print(
"init tracking error, filter_detections len:",
len(camera_port_this),
)
continue
# 满足条件后，初始化全局跟踪状态，添加跟踪目标
global_tracking_state.add_tracking(detections)
tracking_initialized = True
lost_frame_count = 0
# 保留第一帧的3d姿态数据
tracking_initialized = True # 标记已初始化
lost_frame_count = 0 # 丢失帧计数器归零
# 保留第一帧的3d姿态数据，按id存储到all_3d_kps字典
for element_tracking in global_tracking_state.trackings.values():
if str(element_tracking.id) not in all_3d_kps.keys():
all_3d_kps[str(element_tracking.id)] = [
element_tracking.state.keypoints.tolist()
]
print("init tracking:", global_tracking_state.trackings.values())
continue
continue # 跳过本帧后续处理，进入下一帧
# tracking_initialized = True
# ========== 丢失目标处理 ==========
if len(detections) == 0:
# 当前帧没有检测到目标，进入丢失计数逻辑
print("no detections in this frame, continue")
lost_frame_count += 1
lost_frame_count += 1 # 丢失帧数+1
# 进一步完善退出条件：
# 1. 连续丢失阈值帧后才退出
# 2. 若丢失时，最后一次检测到的时间与当前帧时间间隔超过1秒，才彻底退出
@ -1022,18 +1032,19 @@ while True:
last_tracking = list(global_tracking_state.trackings.values())[0]
if lost_frame_count >= lost_frame_threshold:
should_remove = True
# 可选：可加时间间隔判定
# 可选：可加时间间隔判定（可扩展）
if should_remove:
global_tracking_state._trackings.clear()
tracking_initialized = False
global_tracking_state._trackings.clear() # 清空所有跟踪目标
tracking_initialized = False # 允许后续重新初始化
print(
f"tracking lost after {lost_frame_count} frames, reset tracking state"
)
lost_frame_count = 0
continue
lost_frame_count = 0 # 丢失计数归零
continue # 跳过本帧后续处理
# 有检测，正常跟踪
lost_frame_count = 0
# ========== 正常跟踪流程 ==========
lost_frame_count = 0 # 检测到目标，丢失计数归零
# 计算当前帧所有跟踪目标与检测目标的相似度矩阵（多机位）
affinities: dict[str, AffinityResult] = calculate_affinity_matrix(
trackings,
detections,
@ -1044,13 +1055,13 @@ while True:
lambda_a=LAMBDA_A,
)
for element_tracking in trackings:
tracking_detection = []
temp_matrix = []
tracking_detection = [] # 存储每个跟踪目标在各相机下最优匹配的检测
temp_matrix = [] # 打印用：每个相机的最大相似度
for camera_name in affinities.keys():
camera_matrix = jnp.array(affinities[camera_name].matrix).flatten()
detection_index = jnp.argmax(camera_matrix).item()
detection_index = jnp.argmax(camera_matrix).item() # 取最大相似度的检测索引
if isnan(camera_matrix[detection_index].item()):
breakpoint()
breakpoint() # 出现异常时调试
temp_matrix.append(
f"{camera_name} : {camera_matrix[detection_index].item()}"
)
@ -1060,8 +1071,10 @@ while True:
affinities[camera_name].detections[detection_index]
)
print("affinities matrix:", temp_matrix)
# 只有匹配到足够多的检测目标时才更新跟踪（如多于2个相机）
if len(tracking_detection) > 2:
update_tracking(element_tracking, tracking_detection)
# 记录每一帧的3d关键点结果
all_3d_kps[str(element_tracking.id)].append(
element_tracking.state.keypoints.tolist()
)