Research On Cross-camera Pedestrian Tracking Algorithm Based On Spatiotemporal Information

Multi-camera pedestrian tracking aims to determine the location of pedestrians from video streams captured by multiple cameras.In recent years,multi-camera pedestrian tracking has played a huge role in the field of urban public safety.But currently many mainstream multi-camera pedestrian tracking algorithms mainly use the appearance features of pedestrians when linking pedestrian trajectories under different cameras.About the spatiotemporal information is used less.Therefore,this thesis focuses on the multi-camera pedestrian tracking algorithm based on spatiotemporal information.The main work of this thesis is as follows:.(1)When encountering inevitable interaction scenes in real life such as object occlusion and pedestrian interleaving,the probability of pedestrian tracking interruption will be greatly increased.Aiming at this problem,this thesis proposes a single camera pedestrian tracking algorithm based on spatiotemporal correlation.The algorithm enhances the data association between pedestrians by calculating the appearance and spatial-temporal correlation of pedestrians,and designs a hierarchical inference structure with overlapping sliding windows.Pedestrian short trajectories are generated by using correlation clustering algorithms in the window.The use of overlapping sliding windows ensures that the short pedestrian trajectories can be connected into a single camera to complete the pedestrian trajectory.This algorithm can effectively solve the problem of pedestrian tracking interruption and reduce the computational complexity.(2)For the multi-camera pedestrian tracking data association problem,many algorithms use person re-recognition method to solve,but it ignore the pedestrian movement between cameras may have a certain spatiotemporal distribution.In view of this,this thesis proposes a multi-camera pedestrian tracking algorithm based on spatiotemporal constraints.This algorithm constructs a pedestrian spatiotemporal distribution model based on the spatiotemporal information of pedestrians and designs a metric function that integrates spatiotemporal models and visual similarity.This function can Effectively judge whether the pedestrian trajectories under two single cameras belong to the same pedestrian.This algorithm can achieve good nulti-camera pedestrian tracking effect.The method proposed in this thesis was tested on the Duke MTMC multi-target multicamera pedestrian tracking dataset.The performance indicators are better than the comparison algorithm,which proves that the method proposed in this thesis can effectively carry out multi-camera pedestrian tracking tasks.