Joint Views Multiple Human Tracking Based On Spatial Distribution

In recent years,the mobile cameras,e.g.,drone-mounted cameras in the air and wearable cameras on the ground,are able to observe targets in the scene from a flexible perspective.Specifically,top-view cameras can monitor the targets with global and broad field of view,while the horizontal-view cameras can monitor the targets with an immersive and close view.Collaborative analysis of different-view data can facilitate various kinds of applications,such as human tracking,person identification,and human activity recognition.However,top-view cameras can cover more human targets over a high altitude without much occlusion,but have difficulty in discriminating different human targets.Horizontal-view cameras can capture the whole body and more appearance of the hu-man targets but suffer from serious mutual occlusions.For collaborative view analysis,the first step is to associate targets across these two views.This is a very challenging problem due to inconsistent appearance and motion features between top and horizontal views.To this end,this thesis firstly proposes a new approach by matching the targets'spatial distributions between the two views.In particular,a matching cost is defined to decide the actual location of horizontal-view camera and its view angle in the top-view image.In addition,based on multiple human association between top and horizontal views,the thesis integrates the multiple human tracking over frames in each view.And the the-sis proposes the joint-view multiple human tracking based on a dual Markov Decision Process(Dual-MDP)mechanism over the targets' lifetime.Experiments demonstrate the robustness of the proposed methods on the collected joint-view video dataset and the experimental results show the effectiveness of the proposed method.