Fast Target Matching Algorithm Between Non-overlapping Multi-camera

With the continuous increasing demand for monitoring,single camera cannot provide enough monitoring scope.Surveillance systems having more and more cameras develop quickly,that makes the systems increasingly complex.As a result,non-overlapping multi camera surveillance systems are desired to simplify the monitoring system and decrease the cost.The blind spot in the multi camera surveillance systems make the video intelligence analysis quite difficult between the camera monitoring scopes.Consequently,the invetigations of target matching technology for the non-overlapping multi camera surveillance systems have been a hot point,which makes the wide application of surveillance system possible.Target matching technology in non-overlapping multi-camera surveillance system includes feature extraction technique,feature fusion technique and similarity measurement technique.These three parts are the research emphases of this paper.Besides,foreground detection technique is also an important part in this paper as it significantly affects the accuracy and operational efficiency of the target matching technology.This paper investigated the key problems in each part of the target matching technology,and provided effective solutions.Firstly,this paper reviewed the current situation of the three key techniques in the target matching technology at home and abroad,and introduced the classical algorithms for the three key techniques substantively.Secondly,the applicability of classical algorithms to current application scenarios was analysed.Lastly,provided the solution to the problem of low resolution and small target,poprosed the fast target matching algorithm for non-overlapping multi camera surveillance systems,and qualitatively and quantitatively analysed and verified the algorithm.Some creative work has been carriedout as follows:1.This paper proposed a target main color extraction algorithm based on online clustering and a feature sparse optimization method.Since the color distribution in the color images is unknown,manually presetting the final cluster number is unscientific when color clustering the image.In addition,the clustering result is very sensitive about the cluster number.An online clustering algorithm is proposed here,which doesn't need set the cluster number of clusters,and the algorithm can automatically determines the number of the final clusters based on the color distribution of images.Since the feature is sparse after clustering,which will affect the similarity measure results,a feature sparse optimization method is also proposed.It shows that the algorithm can not only effectively improve the clustering results and increase the clustering speed,but also effectively improve the similarity measure results by compared with experimental results.2.The fusion algorithm based on main color features and main color distribution feature is proposed in the study.Since a single feature is not enough for the discrimination accuracy to different targets,this paper proposed a fusion algorithm for the problem,which can improve the discrimination between the different objectives by fusing two features of target.It shows that the algorithm can greatly improve the accuracy of the target matching by compared with experimental results.3.This paper proposed a pixel-level target search strategy based on the target main color feature.Since the target in the scene doesn't always move,foreground detection method based on motion information will fail when the object is stationary.The paper proposed a search method based on main color feature for the problem,wchich can successufully search the stationary object without the motion information.The pixel-level target search strategy is proposed to solve the problem that the search strategies by sliding window have very low operating efficiency and is difficult to achieve real-time status.Experimental results show that the algorithm can not only improve the effectiveness of the search target,but also achieve real-time processing.