A Survey on Multi-Target Multi-Camera Tracking Methods
Multi-Target Single Camera Tracking limits the target tracking range to the field of view of a single camera,which is difficult to meet the needs of complex application scenarios.Multi-Target Multi-Camera Tracking(MTMCT)fuses the information from multiple cameras to realize the feature transfer and trajectory correlation between multiple cameras,which can jointly track multiple targets across multiple cameras in multiple monitoring areas,which is of great significance to the real-time monitoring of real-time complex scenarios.It has become a research hotspot in the field of Multi-Object Tracking.This paper introduces the basic concepts of MTMCT and classifies the tracking models into three categories,including overlapping view,non-overlapping view and mixed view,and the MTMCT model with an overlapping view is analyzed and introduced in detail in five subclasses:network flow optimization-based method,homography constraints method,reinforcement learning method,hypergraph method and transformer method.subclasses are analyzed and introduced in detail.The network flow optimization method is a special Bayesian network model to solve the MTMCT task by calculating probabilities from binary images detected by multiple cameras.Methods based on network flow optimization can cause the phenomenon of false targets to affect the detection results,so the homography constraint method appears.The homography constraint defaults to all targets being of the same height,and the detection results of all views are projected onto the same ground plane after adding the constraint,and the intersection point formed is the position of the target.With the continuous development of deep learning methods,methods based on reinforcement learning appear,which mainly use hierarchical combination models to track the target.Most of the reinforcement learning methods are offline,and the algorithm parameters are large and difficult to train.In order to solve this problem,hypergraph-based methods have emerged,which introduce weighted hypernet work into the MTMCT task,and the weighted hypernet work can reduce the number of parameters and redundancy by sharing parameters across layers.The current transformer-based method can be realized to achieve online tracking with good performance at the same time.MTMCT with non-overlapping viewpoints is divided into two types of two-stage trajectory association and single-stage trajectory association for detailed analysis and introduction.MTMCT based on two-stage trajectory association refers to outputting multi-target tracking trajectories within a single camera first,and then carrying out multi-camera trajectory associations;MTMCT based on single-stage trajectory association refers to directly considering all trajectories globally to be associated.Different from the overlapping and non-overlapping views,the mixed view MTMCT method can be used on both overlapping and non-overlapping view datasets with a good balance of algorithmic performance and accuracy.In addition,we also compare the advantages and disadvantages of overlapping view,non-overlapping view,and mixed view approaches and their applicability scenarios.Finally,this paper analyzes the commonly used datasets and evaluation indexes for MTMCT,and summarizes the problems of MTMCT.We also look forward to the future trends of MTMCT technology,such as more MTMCT datasets,end-to-end models,richer evaluation metrics,higher crowd density,visual Transformer,and lightweight models.
Multi-Target Multi-Camera TrackingMulti-Object Trackingcamera association modeloverlapping perspectivenon-overlapping perspectivemixed perspective
万方数据
维普
