Overview: This paper model the multi object tracking process with a MDP to fully utilize the detector; All objects in each frame is in Active, Inactive, Lost and Tracked 4 states; And the system needs to come up with the decision of transition (i.e. policy; it could be deterministic or with learnable parameters).
Tracking mode: Batch means to take in all frames at once and has the visibility of the future frames; Online means the visibility up to the current frame.
Active -> Tracked/Inactive: A binary SVM will classify all detected objects into tracked or Inactive depending on the coordinate, height, width and detection score.
Tracked -> Tracked/Lost: An optical flow from sample points in the detected bounding box will be calculated and use FB (forward-backward) error to measure how stable the motion is; To filter out background false alarm, A small displacement is required.
Lost -> Tracked or Lost or Inactive: A new detached objects will be compared with all possible association with lost sample’s certain history. If it stays lost for long enough, it because inactive.
Inverse RL: The author mentioned this systems needs to learn the reward function from the training data. The method is mentioned in one of Andrew’s early year paper ‘A. Y. Ng and S. J. Russell. Algorithms for inverse reinforcement learning. In ICML, pages 663–670, 2000. 3
’ I need to check this out.
Evaluation metrics: This paper mentioned some evaluation metrics for MOT in section 4 that capture most of the aspects we cared.