Research Of Multi-camera Object Tracking Based On Non-overlapping Views

In recent years, with the rapid development of technology in video surveillance, and the increasing demand for intelligent monitoring, the traditional single-camera monitoring system has failed to meet the requirement of video surveillance on many situations. Compared with single-camera monitoring system, using multiple cameras can expand observation range, increase observation angle, capture more information of moving objects, and deal with the problem of occlusions. Therefore, it has been attracted more and more attentions. It often does not have overlapping views between adjacent cameras on many occasions, especially when multi-camera system monitors a large range of occasions. Therefore, multi-camera video surveillance based on non-overlapping views becomes a direction for research with practical value.In the environment for multi-camera surveillance based on non-overlapping views, 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, affecting the accuracy of objects tracking and matching. To deal with the problem, tracking and matching objects across non-overlapping camera views is studied in this thesis. The main contributions and innovation points are as follows:(1) Studies on the establishment of brightness transfer function (BTF), and an effective method for adjusting BTF is proposed according to scene brightness of camera views. Firstly, we judge scene brightness between camera views to determine the direction of mapping. Then, the establishment of the BTF is performed according to the normalized cumulative histogram of objects. This method can avoid the loss of some brightness mapping, and we verify the necessity for performing BTF.(2) The properties of mean brightness transfer function (MBTF) and cumulative brightness transfer function (CBTF) is analysed. On the basis of the above, we combine the idea of both multi-frame and multi-object, and propose a synthetical BTF. The modified Bhattacharyya coefficient is introduced as well to measure the accuracy of the brightness mapping. Experimental results show that the synthetical BTF performs more accurate for brightness mapping. (3) An objects correspondence algorithm based on multi-frame matching which uses probabilistic principal component analysis (PPCA) to deal with BTF is proposed. The collection of brightness transfer functions is obtained by training keyframes of several pairs of objects, and PPCA is introduced to probability estimate the space of brightness transfer functions, adopting multi-frame matching to calculate the probability of BTF between objects belong to the space of brightness transfer functions. Simultaneously, considering the lack of training objects, combined with the method of the synthetical BTF, adopting multi-frame matching as well to calculate the similarity between objects. Experimental results show that this method performs well for objects correspondence.Finally, the summary of this thesis is made, and the future work is discussed based on the work of this study.