Research On Multiple Object Tracking Method Based On Joint Re-detection And Re-identification

As an indispensable core technology in intelligent video surveillance,automatic driving and other applications,multiple object tracking(MOT)has become the most popular research topic in the field of computer vision.At present,the mainstream MOT models adopt detection-based tracking mode,under this mode,limited by the shortcomings of existing object detectors and the complexity of MOT scenarios,missing detection severely limits the development of MOT,and has become an urgent problem to be solved.On the one hand,most of the existing multi-object trackers believe that the missing detection is only brought by the detector,and usually uses the motion model to predict the position of the known trajectory in the current frame,and uses it as a supplement to the detection result to solve the missing detection.However,the performance of the motion model is not outstanding when the motion state of the object changes greatly.On the other hand,the application of non-maximum suppression(NMS)algorithm in MOT is very common.When the intersection over union(Io U)of two candidate bounding boxes(BBoxes)reaches the threshold,it will delete the one with lower confidence score.However,in crowded scenes,even two BBoxes with different identities will have a large Io U due to occlusion,and NMS will mistakenly delete one of the objects,resulting in new missing detection.The missing detection caused by NMS has never been paid attention to in the existing MOT trackers,and this thesis proposes a solution to this missing detection for the first time.The main research work of this thesis is as follows:(1)For the missing detection brought by the detector,this thesis designs a trajectory re-detection(T-Re Det)network based on the Siamese network architecture,which predicts the position of the trajectory in the current frame by detection and generates more candidate BBoxes to avoid missing detection.Specifically,T-Re Det network first uses the Siamese network to extract features for the trajectory and its corresponding image crops in the current frame,and then finds the most likely position of the object center in the current frame through the depth-wise cross correlation layer.Finally,the BBox of the trajectory is regressed by the anchor-free detection method.T-Re Det network is essentially a lightweight detection model for a specific object based on local region feature matching.Compared with the prediction method based on the motion model,T-Re Det network is less affected by the change of the object motion state and is more reliable.(2)For the missing detection caused by NMS,this thesis proposes to integrate the re-identification technology into the standard NMS(Re-ID NMS).A double check is formed by combining the appearance similarity calculated according to the re-identification feature and the spatial coincidence expressed by Io U,so as to jointly judge whether the two BBoxes represent the same object.Compared with the standard NMS that only relies on Io U to filter out redundant candidates,Re-ID NMS can filter out redundant candidates more accurately keeps candidates that should have been deleted directly.(3)In this thesis,the proposed T-Re Det network and Re-ID NMS are applied to the latest MOT tracker TADAM,and a more perfect tracker is constructed.Extensive experiments on three datasets,MOT16,MOT17,and MOT20,fully demonstrate the effectiveness of the method proposed in this thesis.