Slightly improved tracker.

rpt/tracker.hpp

@@ -156,7 +156,7 @@ std::vector<std::tuple<size_t, std::vector<std::array<float, 4>>>> PoseTracker::
         }
     }
 
-    // Remove old tracks
+    // Remove old track entries
     timestamps.push_back(timestamp);
     if (timestamps.size() > history_size)
     {
@@ -166,8 +166,21 @@ std::vector<std::tuple<size_t, std::vector<std::array<float, 4>>>> PoseTracker::
     for (size_t i = 0; i < pose_tracks.size();)
     {
         auto &track = pose_tracks[i];
-        double last_timestamp = track.timestamps.back();
-        if (last_timestamp < max_age)
+        for (size_t j = 0; j < track.timestamps.size();)
+        {
+            double ts = track.timestamps[j];
+            if (ts < max_age)
+            {
+                track.core_poses.erase(track.core_poses.begin() + j);
+                track.full_poses.erase(track.full_poses.begin() + j);
+                track.timestamps.erase(track.timestamps.begin() + j);
+            }
+            else
+            {
+                j++;
+            }
+        }
+        if (track.timestamps.size() == 0)
         {
             pose_tracks.erase(pose_tracks.begin() + i);
         }
@@ -177,23 +190,15 @@ std::vector<std::tuple<size_t, std::vector<std::array<float, 4>>>> PoseTracker::
         }
     }
 
-    // Remove old poses from tracks
-    for (auto &track : pose_tracks)
-    {
-        while (track.core_poses.size() > history_size)
-        {
-            track.core_poses.erase(track.core_poses.begin());
-            track.full_poses.erase(track.full_poses.begin());
-            track.timestamps.erase(track.timestamps.begin());
-        }
-    }
-
     // Refine poses
     std::vector<std::tuple<size_t, std::vector<std::array<float, 4>>>> tracked_poses;
     for (size_t i = 0; i < pose_tracks.size(); ++i)
     {
         auto &track = pose_tracks[i];
-        if (track.core_poses.size() > 0)
+        // Create a refined pose for current tracks, or old tracks with a bit history,
+        // to avoid continuing tracks of flickering persons
+        if (track.core_poses.size() >= std::ceil(history_size / 2.0) ||
+            track.timestamps.back() == timestamps.back())
         {
             std::vector<std::array<float, 4>> refined_pose = refine_pose(track);
             tracked_poses.emplace_back(track.id, refined_pose);
@@ -205,7 +210,8 @@ std::vector<std::tuple<size_t, std::vector<std::array<float, 4>>>> PoseTracker::
 
 // =================================================================================================
 
-std::tuple<int, float> PoseTracker::match_to_track(const std::vector<std::array<float, 4>> &core_pose_3d)
+std::tuple<int, float> PoseTracker::match_to_track(
+    const std::vector<std::array<float, 4>> &core_pose_3d)
 {
     int best_track = -1;
     float best_distance_sq = max_distance * max_distance;

scripts/test_skelda_dataset.py

@@ -128,6 +128,8 @@ datasets = {
         "path": "/datasets/chi3d/skelda/all.json",
         "take_interval": 5,
         "fps": 50,
+        "min_match_score": 0.92,
+        "min_bbox_area": 0.2 * 0.2,
         "max_track_distance": 0.3 + default_max_movement_speed / (50 / 5),
     },
     "tsinghua": {
@@ -193,7 +195,7 @@ datasets = {
         "subset": "volleyball",
         "min_match_score": 0.95,
         "min_group_size": 7,
-        "min_bbox_score": 0.25,
+        "min_bbox_score": 0.20,
         "min_bbox_area": 0.05 * 0.05,
         "max_movement_speed": 4.0 * 1.5,
         "max_track_distance": 0.3 + (4.0 * 1.5) / (20 / 2),