Research On Object Tracking With Non-overlapping Views Based On Appearance Features And Spatio-temporal Relationships

Tracking across cameras with non-overlapping views is one of the most challenging tasks of intelligent surveillance systems. Unlike the closely related task of single-camera tracking, disjoint inter-camera tracking is difficult due to two aspects of reasons. Firstly, the observations from non-overlapping views are widely separated in time and space. Secondly, the appearance of an object in one camera view might be very different from its appearance in another camera view due to the differences in illumination, pose and camera properties, which may affect the accuracy of target tracking and matching.To deal with the problems of tracking across cameras with non-overlapping views, both the appearance and space-time relationships are taken into consideration. The main contributions and innovation points are as follows.(1) Multi feature based trajectories reconstruction and features extraction. It’s difficult to reconstruct the image point to world coordinate system when each scene is only corresponded to one camera even though it is fully calibrated. The proposed method firstly extracts the foot and head points of targets by using principal axis, forming a constraint equation. Secondly, trajectory smoothness term is added as an additional constraint equation. Finally, though the projection equation, the optimization equations are established to reconstruct the foot and head point in world coordinate system. The velocity and orientation features are extracted. And the spatio-temporal correspondence probability is confirmed by trajectories’temporal correlation.(2) Estimation of object’s height by multi plane foreground pixels projection. An appearance based model using3D information is created for each object, encoding appearance information and temporal information together. It also provides an easy approach to refine and update the model if multiple shots from multiple cameras are available.(3) Finally, with the establishment of a probabilistic model, the association probability of the space-time relationship and that of the appearance based matching are combined together. The correspondences problem is solved by finding the maximum likelihood probability solution which can be optimally solved by finding a maximum match of a bipartite graph。The experimental results show that the effectiveness of the proposed algorithm.