An Video Moving Objects Segmentation Algorithm Based On SVM And GMM

Moving objects segmentation and detection is not only one of the most important issues in the digital video processing and computer vision researching areas, but also the key technology of intelligent video surveillance applications. Even though people have done a lot of work related with video objects segmentation, some inadequacies still preserved. There is not an all-purpose method which would be able to effectively separate moving objects from the background, so the segmentation of moving objects from the video sequence has become the hottest research currently. Gaussian mixture probability model based on background subtraction is the classic algorithm, with many advantages such as easy implement, the background modeling with multi-peak distribution as well as the self-adaptability. But the complex background (the light changes, the leaves sway and the rain and snow) impact on the object segmentation effect which greatly restrict its application.Paper introduces the development of video object segmentation technology, analyses the video object segmentation theory, and then make a in-depth study of Gaussian mixture models and support vector machine theories, and their approaches of video object segmentation. On that basis, a novel algorithm which integrates Gaussian mixture model with the support vector machine classifier is proposed. The algorithm is characterized by making full use of the advantages of the traditional Gaussian mixture model at the same time making up for deficiency of model building which just use space-domain information. The support vector machine based on statistical learning theory to make segmentation between foreground and background improves object segmentation quality and stability. This method firstly made preliminary segmentation to get binary image though GMM. Then the pixels in every frame were classified as background and foreground pixels which composed the corresponding background and foreground blocks. In order to obtain the time-domain information, the representative three frames (the background frame, the current frame, the previous frame of current frame) had been partitioned into sub-blocks, according to the corresponding blocks in these frames statistics elements were obtained as eigenvectors. Then, we trained SVM classifier with feature vectors, completed selection of kernel function and optimization parameters, and used the trained classifier to classify pixels which were constituted into the partition templates. At last, the integration of SVM and GMM can be given simply by the intersection operation of segmentation results.Comparing the experimental results of proposed algorithm with the inter-frame difference method, approximate median algorithm and Gaussian mixture model algorithm, from the results of both subjective and objective evaluation, experimental results showed that the proposed approach significantly decreases the false motion detection and improves segmentation quality of moving objects. The algorithm is applicable, in particular, suitable for outdoor video as well as the road traffic intelligent monitoring with complex background.