Study Of Target Handover Algorithm Based On Spatial-Temporal Transfer Model

Multi-camera target handover is not only a key step of large-scale monitoring but also a research focus in computer vision. In this paper, the three main issues of target handover including objects detection, feature representation and target matching are studied.Firstly, moving target detection under different background is studied. The double difference method is applied to detect moving target under static background. It can not only improve accuracy but also ensure the real-timing. The fast block matching method based on K-means clustering is researched to overcome the background changing in target detection under moving background. The global motion vector and local motion vector are obtained by block-matching method, and then the foreground targets are segmented by K-means clustering. This method can effectively detect the moving target under dynamic background.Secondly, the appearance model and spatial-temporal model of the target are built. According to the feature difference, color characteristics and SIFT features are selected to constitute the appearance model of the target. Brightness transfer function is also used to correct color of target under different cameras. The spatial-temporal model consists of spatial transition probability and temporal transition probability. Spatial transition probability is obtained by moving orientation estimation of target and camera topology, while temporal transition probability is estimated by moving velocity of target and distance between cameras.Lastly, the spatial-temporal model and the appearance model are combined under Markov chain to handover. The hardware and software configuration of the test platform are briefly introduced. The algorithm based on the spatial-temporal model is applied to experiment on the platform under over-lapping view field and non-overlapping view field. The experimental results show that the handover algorithm based on spatial-temporal model is effective.