Moving Area Detection And Tracking Based Video Condensation

With huge amount of cameras is being deployed in cities, the endless amount of surveillance video data have caused reviewing those video to be a very difficult task. In order to use huge processing capacity of modern computer system to condense these huge videos, a novel method, based on moving area detection and tracking, to condense the surveillance video data is proposed. The proposed algorithm firstly constructs the background using mix Gaussian model, segments foreground area of the video stream according to both the color and gradient differences of the background and foreground. Mean-shift based target tracking algorithm is used to trace moving areas, as result of which, continuous masks of activities representing moving targets are collected and stored. Activities’SIFT vector and moving trace distances between each other of these slices are computed and indexed. Furthermore, based on the weighed sum of SIFT vector distance and moving trace distance spectral clustering method is used to cluster these slices. Finally, these clustered slices are combined together according to their space position in the original video. Users could also search similar activities of one specified slice or activities using Nearest Neighborhood method. As a result, the method could automatically extracts activities from the huge amount of video data, analysis the features of these activities to classify them, shows these clusters separately in form of new condensed video which makes storing and checking surveillance video more convenient. A MFC based framework is implemented to testify the method talked above. Background construction foreground segmentation and activity extraction algorithms are implemented in the system. Based on the Object-Oriented soft engineering techniques and various design patterns, the system could be easily extended with more algorithms and be re-developed by any third-party.