Research On Multi-camera Target Detection And Tracking Algorithm For Non-overlapping Domains

With the expansion of the intelligent video surveillance area, traditional single-camera target detection and tracking technology cannot meet the monitoring demand anymore, therefore, multi-camera detection and tracking technology need to be studied. As an important subject of intelligent video surveillance research area, multi-camera detection and tracking has a strong theoretical significance and application value. Researches on multi-camera detection and tracking technology can be classified into two scenarios: overlapping domain and non-overlapping domain. Because the surveillance range in reality is wide, most monitoring frames are none-overlapped, in other words, there exists blind spots. Therefore, research contents of this paper focus mainly on multi-camera detection and tracking of moving targets in none-overlapping domain.Contents of this thesis are as follows:An improved ViBe algorithm for single-camera target detection is proposed, which combines consecutive three- frame differential algorithm and traditional ViBe algorithm, to roughly extract target active area for updating background classification. Besides, for the purpose of foreground image integrity insurance, the proposed algorithm fills holes by means of morphological processing methodMoving target feature models are composed of the color histogram, multi- scale LBP feature and SURF feature. For the problems of the illumination discrepancy from different camera views, luminance transfer function is implemented for illumination compensation. By using characteristic weighting method, Multi scale LBP operator is used to solve the problems of low classification performance and incomplete description of single scale LBP operator. For to solve problems of target image differences in each camera, SURF feature pairs is matched by using Euclidean distance ratio. Matching pairs is purified by using improved RANdom Sample Consensus(RANSAC) algorithm, which solves the mistaken paired problem.In order to reduce the number of feature matching, it is proposed that multi camera temporal constraint relation is used to match the constraint information. Spatio-temporal constraint relation includes spatial transition relation and tempora l transition relation.Using D-S evidence theory to fuse multi- features and spatio-temporal constraint relation, D-S evidence theory based target relay method is proposed, which avoids the relay errors of single feature to improve the robustness of the system.