Multi-Target Multi-Camera Tracking With Face Joint Identification And Global Trajectory Pattern Consistency

With the rapid development of modern cities,the density of urban population has exploded exponentially.At the same time,the safety monitoring and automated analysis of urban crowd behavior has become an important topic in computer vision,which is of great significance for preventing and tracking major public safety accidents.Multi-Target Multi-Camera Tracking is also receiving more and more attention from researchers.The goal of Multi-Target Multi-Camera Tracking is to determine as much as possible the trajectory of all pedestrian targets under all cameras to monitor and analyze crowd behavior.Specifically,Multi-Target Multi-Camera Tracking requires tracking pedestrians under each camera as much as possible and accurately,and then merges these trajectories with representation learning and clustering algorithms.Finally the trajectory of each pedestrain across all cameras can be determined.At present,Multi-Target Multi-Camera Tracking algorithm based on deep features has achieved the considerable performance.Compared with traditional handcraft features,deep features have richer semantic information and stronger representation capabilities.In this paper,we first investigate the specific implementation details of the existing algorithms,analyze their advantages and disadvantages,and then improve them.We propose our own solutions,and verify them through experiments.Recent Multi-Target Multi-Camera Tracking algorithms are mainly based on the body apparent features,using deeper networks and more complicate loss functions to learn more representative features.They use some complex clustering algorithms when tracking across cameras and have achieved good results on public benchmark datasets.However,in real scenes,body features are often not enough to deal with all situations,and too complex clustering algorithms lead to inefficiency,which directly limits the practical applications.In this paper,we make improvements based on existing methods.we build a face-body joint tracking and recognition framework,combining face and body appearance features to identify pedestrian targets.In order to ensure the efficiency of the algorithm,a more lightweight clustering algorithm is used in this paper.What's more,we adopt a temporal-spatial constraint named global trajectory consistency to further regularize the results.The results on the public benchmark datasets show that,the proposed algorithm achieves a great performance improvement and achieves state-of-the-art comparing with the algorithms published so far.