Research On Multi-Camera Multi-Target Tracking

As surveillance video networks become more and more complex,video analysis under a single camera is insufficient to meet people's needs.For multi-camera multi-target tracking technology,since it can effectively realize automatic monitoring in a wide range of fields of view,it received a lot of attention.The main task of multi-camera multi-target tracking is to obtain the complete trajectory of the target under multiple cameras.This thesis mainly studies some key issues involved in multi-camera multi-target tracking from the following aspects.First of all,because the traditional multi-camera multi-target tracking method mostly adopts the "two-step" approach: first,single-camera multi-target tracking and then multitarget tracking across cameras,there is an error in single-camera multi-target tracking.This error will be further amplified during the multi-target tracking phase of the camera.In response to this problem,this thesis adopts a "one-step" approach,which combines singlecamera multi-target tracking and cross-camera multi-target tracking for global optimization.By placing single-camera multi-target tracking and cross-camera multi-target tracking into a global graph model,global trajectory fusion is performed to avoid as much as possible when multi-camera multi-target tracking is adopted by the "two-step" approach.The wrong match and the miss match problem.Secondly,the introduction of global ideas into multi-camera multi-target tracking,although it can bring the advantages mentioned above,yet it introduces a new problem,that is,when we perform global optimization,The distribution of apparent feature similarity across cameras will be much lower than that under a single camera.If the same similarity measure is simply used to measure the similarity of any two tracks,it is highly likely that in the optimization process,the global map model preferentially connects the trajectories in a single camera,ignoring the trajectory connection across the cameras,thereby affecting the overall effect of multi-camera multi-target tracking.Aiming at this problem,this thesis proposes a multi-camera multi-target tracking algorithm based on the equalization global graph model.By equalizing the distribution of these two similarities,it is guaranteed that the two distributions in the global graph model optimization process the balance,to some extent,avoids the above mentioned problems and improves the overall effect of multicamera multi-target tracking.Finally,using the results obtained by the above two algorithms,this thesis proposes a multi-camera multi-target tracking algorithm based on space-time constraints.On the one hand,the spatial constraint relationship of any two trajectories is calculated by proposing a relative distance concept.On the other hand,the time constraint relationship is introduced by estimating the transition time probability distribution between the camera network topological nodes.In the multi-camera multi-target tracking,by introducing the space-time constraint relationship,on the one hand,the computational complexity of the target association can be reduced,and on the other hand,the accuracy of the target association is also improved.The experimental results show that the proposed method achieves the best results in the multi-camera multi-target tracking evaluation indicators,which fully proves the robustness and effectiveness of the model.