Based On The Target Under Binocular Visual Motion Detection

Using binocular vision technology to deal with video is the forefront of computer vision research field today, and is widely used in the fields of military and civilian. This paper has studied mainly on stereo matching under binocular vision, global motion estimation and objects detection under moving background based on the binocular vision technology, for the final purpose of fulfilling objects detection under moving background.This paper uses Markov Random Field (MRF) model for stereo matching. It introduces a data function that integrates gray with gradient to build a MRF global energy function, and uses improved belief propagation to minimize the global energy function in order to get a disparity map, which is then calibrated by MRF causal systems. Experiment results show that accuracy of stereo matching has been improved by this method.Global motion estimation is a key step in objects detection under moving background. We proposed a feature point-based global motion estimation algorithm which combines least median of squares(LMedS) and least-squares(LS) methods. First, we build a motion model according to the global motion, extract feature points using SUSAN algorithm, and then match these points using a weighted circular template in conjunction with the similarity measure that combines gray and SUSAN initial response values. At last, we get parameters of global motion estimation using least median of LMedS and LS methods. Experiment resultsshow that the algorithm is effective.The objects detection, that uses the background subtraction algorithm under staticbackground based on binocular vision, is effective in cases that the ambient light changes suddenly and the objects have a similar gray with the background. This paper completes the objects detection under moving background using gray inter-frame difference method based on single camera, then proposes an objects detection method combining gray and disparity inter-frame difference under moving background based on binocular vision. This method is more robust than the one using only gray or disparity. When several objects have an overlap area, the objects segmentation can be effectively performed by the disparity map.