Research On Video Multi-target Association With Online Learning

As the monitor plays an increasingly important role in surveillance in maintaining social stability, the number of surveillance video shows an "explosive" growth. In the face of massive video data, multi-target association and tracking method faces challenges in computational complexity and real-time demand etc. Association and tracking system is a dynamic estimation problem, using the continuity of target in time and space correlation. However, as the video data is incomplete caused by damage or target association across different cameras and videos is needed, the target associated is no longer continuous in space and time. Thus, association methods based on the video time clues and target motion information between successive frames may fail. In view of the problems above, a video multi-target association technology is proposed which is based on online learning theory. The technology includes several aspects listed as follow:(1) Online learning method is proposed for large-scale data processing problem. Online learning is an effective approach to solve training complexity matter of large-scale data. Video data arrives in stream which conforms to the training data characteristics of online learning. Each arrival invokes an update on the current model rather than waiting until all the data obtained which improves the operational efficiency. Moreover, the real-time updated classifier can be better adapted to the change in data characteristics.(2) Design and implement the online support vector machine and its predictor sparsity. Kernel-based SVM algorithm is outstanding in dealing with the problem of nonlinear classification. The paper turns it into an online form using Fenchel duality of convex optimization theory, and proposes a classifier updating mechanism of window function method through the analysis of dual ascending. To selectively update the classifier, improves the sparsity of online model and to constrain the dual ascending, enhances the robustness of the classifier to noises, thus ensure the accuracy of data classification. Experimental results on multiple datasets verify the effectiveness of the algorithm.(3) Design and implement the multi-target association technique in reference to multi-target classification. This technique is based on the classifier of online support vector machine, and do not need the continuity thread of target in time and space as association clues. So the target can be well associated while the video data is incomplete, missing frame, or even the frame sequence is disrupted. In addition, this target association technology is also applicable for multi-target association across different videos and cameras without having to reconstruct the correlation model.(4) For real-time demand, a parallel detection and association structure is proposed. The two part, detection and association of the multi-target tracking system, is detached as two separate modules for the parallel processing of video. That is, multiple video frames can be processed at a time. This structure can effectively solve the contradiction that the real-time demand and the accuracy cannot be increased at the same time in the general multi-target detection and association method.Multi-target association experiments is conducted on multiple video datasets respectively and the result is analyzed. The experimental result verifies the performance of the method proposed.