Space-time Salient Objects Tracking In Video Surveillance

Video data exists in unstructured form with huge quantity. Saliency detection simulating the fast extraction of information of human vision system has become a hot spot of computer vision. Salient objects are the interesting and major entities of images and videos. Thus detecting locating recognizing and tracking objects with the help of salient object detection can achieve high efficiency in video surveillance.Recent researches show that fusing multiple features can obtain saliency maps with high quality. In this paper, we take video sequences as the research object, and investigate object detection segmentation and tracking of space-time salient objects. The details of our work are as follows:1) Traditional saliency detection methods take frame center as the position prior. The motion continuity will be ignored. Besides, using boundary and connectivity as background prior cannot deal with scenes of complex texture. These problems lead to the incorrectness of salient object extraction. Considering the spatial continuity of salient regions in adjacent frames, we propose a salient object detection method combing improved priors with space-time features. This method uses the result of detection to compute position prior. A mean filter in temporal dimension is used to obtain the background to compute background prior, and then combine these priors with velocity and acceleration to generate the prior map, finally combine the prior map with the feature contrast map generated by color, motion magnitude and orientation to obtain the final saliency map. Experimental results on standard datasets and surveillance show that the proposed method can effectively suppress background and highlight salient objects.2) The determination of salient region is the key of the integrity of salient objection extraction. Focus on this problem of existing methods, we start from saliency maps and exploit the improved region grow method to extract salient objects in saliency maps based on the analysis of algorithms. The experimental results on surveillance videos show that the modified region grow method can achieve good performance to ensure relatively complete object extraction for subsequent object tracking.3) Focus on traditional Mean Shift tacking method cannot the update of the templates, which leads to the failure of tacking the object with quick change, and the problem that Kalman filter and Mean Shift cannot deal with multi-object tracking, we use the combination of Kalman filter and Mean Shift based on salient object detection and data association to track salient objects. This method updates templates of objects in traditional Mean Shift tracking method. The Kalman filter is used to predict the location of objects in the next frame to speed up Mean Shift algorithm. Data association is applied to matching detections and templates. The experimental results show that the proposed method can improve the accuracy of object tracking.