Research On Salient Motion Of Object Detection In Surveillance Video

With the increase of global terrorist activities and the development of monitoring equipment,people pay more attention on video surveillance technology,and automatic detection of suspicious moving targets in the surveillance video is a key point of video surveillance.Currently,there still exist certain difficulties in the process of detection,such as complex motion in the natural scene,the disturbances caused by camera shake,which increase the inaccuracies of motion detection.Distinguishing the salient motion from the disturbed motion is a difficult and major research in the field of moving targets detection in surveillance.By fusing the spatiotemporal saliency,this paper proposes an object detection method based on the salient motion.The main work is as following:In the spatial saliency detection,the weighted digraph is proposed to express the image,the improved SLIC algorithm is used to obtain super-pixel blocks,which are the nodes of graph,the neighborhood distances of super-pixels are the arcs of graph,while the neighborhood information of terminal nodes is taken into account,and then generate the corresponding Hamiltonian matrix.Finally,by calculating the eigenvectors of the minimum eigenvalues,the spatial saliency map of single-scale can be obtained.In the temporal saliency detection,pixel-based sampling background modeling method is proposed.With the continuity of video frames,this method builds the pixel-based sample models and updates it continuously.By this model,motion saliency value of the pixel is calculated,the temporal saliency map can be obtained.In the spatiotemporal saliency detection,the above spatial and temporal saliency maps are fused by the dynamic-coefficient fusion method and the salient motion is detected.In this paper,Pub PETS and GTI,two public datasets were used to test the proposed algorithm in MATLAB coding.Four popular salient motion detection methods,Three-Frame-Difference,Mixture of Gaussian model,CodeBook,and Vibe,are compared with the proposed algorithm by three evaluation indexes of precision,recall and F-measure.Experimental results show that the proposed method has more detection accuracy than the four methods.It is robust for light change and less time-consuming,which can meet the demand of real-time.