Benzlxs / tracking

Tracking

Testing procedures

When downloading the new sequence of vedio.

1. Run ground truth using extract_gt_tracklets.py;
2. Creat the cropped point with function: create_reduced_point_cloud in detections_gt_generation.py;
3. Estimate the road plane parameters using matlab code under folder /dataset/extract_ground_planes ;
4. Generate the detection results with code detections_gt_generation.py;

After generating all the detection results, the tracking visualizaton function should be launched to check all the detectio results are correct.

The evaluation procedures

1. run eval.py with run_det_trk_results function;

Save the detection and tracking results in simulation:

1. save without running the detector again, save_detection_tracking_multi_phases_without_detection in sequence_pc_detection.py;
2. save with running the detector, save_detection_tracking_multi_phases in sequence_pc_detection.py ;

Accuracy in the simulation procedures

1. to generate the function with save_detection_tracking_multi_phases in sequence_pc_detection.py；
2. to evaluate the function with run_det_trk_result_in_simulation in eval.py;

Efficiency analysis in simulation:

1. efficiency_analysis_all_phases() and efficiency_analysis_all_phases_with_fusion_model() in eval.py;

To analyze the efficiency and accuracy results with real tracker

1. regenerating the detection results with the model used for similuation using the code detections_gt_generation.py;
2. generating the tracklet with tracker and save them under xxxx_sync/detection/tracklet_det, with function pointcloud_tracking_classification_with_saving_tracklets() in pointcloud_tracking.py;
3. running tracker to generate tracking results and put them under xxxx_sync/detection/tracklet_trk, with code function fusion_model_with_sort_results in tracker/fusion_model_sort.py;
4. producing the accuracy comparison with the function run_det_trk_result_with_real_tracker() in eval.py;
5. producing the efficiency comparison with the function

File descriptions

1. /detection/gt: bounding boxes extracted directly from tracklets with code: /dataset/extract_gt_tracklets.py save_dets();

2. /detection/dets_gt: bounding boxes generated by segmentation algorithm with code: /dataset/detections_gt_generation.py save_detection_gt(), its xyzlwh will be slightly different from gt results;

3. /detection/dets_class: bounding boxes and confidence are generated by classification model with code: /dataset/detections_gt_generation.py save_detection_classification_multi_phases();

4. /detection/tracklet_det: tracklets with real tracker with code: pointcloud_tracking.py pointcloud_tracking_classification_with_saving_tracklets();

5. /detection/tracklet_trk: tracklets with fused confidence with code: pointcloud_tracking.py fusion_model_with_sort_results();

6. tracklets_pc/seq_id: the points of object in one tracklet with code: /dataset/sequence_pc_extract.py save_sequence_pc_frames();

7. tracklets_pc/dets_conf: detection result for every seq_id with code: /dataset/sequence_pc_detection.py save_detection_tracking_multi_phases();

8. tracklets_pc/trk_conf: tracking fusion result for every seq_id with code: /dataset/sequence_pc_detection.py save_detection_tracking_multi_phases_without_detection();

Detection accuracy in /detection/tracklet_det and tracklets_pc/dets_conf should be the same.