Binocular Vision-based Moving Objects Detection And Tracking

Binocular stereo vision is an important branch of Computer vision. It can get flexible and accurate scene depth information in a variety of conditions. And it has important research value and significance in the field of image processing and computer vision. Moving target detection and tracking has always been the core subject in the field of computer vision. It has a very important practical value and broad development prospects in the military vision-guided, video surveillance, virtual reality, medical diagnostics, intelligent transportation, industrial product testing and other aspects. Traditional method is based on monocular vision, it has difficult in moving object detection and tracking in the complex scene. Especially in the occlusion occurs and multiple moving targets, it is difficult to accurately track the moving target. To solve the above problems, this paper uses binocular stereo vision technology to detect and track moving targets, the main research work are:1.Doing in-depth study of traditional stereo matching algorithms. For the slow, poor accuracy and mismatch problems of traditional stereo matching algorithms, we proposed a fusing optical flow stereo matching algorithm. This method doesn't need epipolar constraint, therefore the range of correspondence research is within a square area centred on the seed point position. Furthermore, the correspondence research is conducted not only in the left and right image pairs, but also in continuous image sequences. So, for the right image sequence, the optic flows can be calculated on all candidate points to be matched with the seed points in the left image. It improves the matching speed and accuracy.2.Through research on the background subtraction. We analyze the advantages and disadvantages of monocular and binocular stereo vision in target detection. Then propose a target detection method which combines disparity-based moving target detection algorithm and intensity-based moving target detection algorithm. This method uses the intersection of foreground targets detected by the two algorithm as moving objects. It overcomes the binocular vision in target detection difficult to accurately obtain the target profile, monocular vision vulnerable to impact of environmental change. And it can detect targets correctly unaffected by changing lighting conditions and the shadow of the object is irradiated and the objects are occluded by each other. Thereby this method enhances the stability of target detection.3.Based on the detected moving targets, to analyze CamShift tracking algorithm, and then for CamShift tracking algorithm is easy to interfere by similar colors of the background and target problems,we proposed a combination of CamShift tracking algorithm and binocular stereo vision tracking algorithm to achieve the tracking of moving targets. In this method, we use disparity-based background subtraction to detect moving targets, and get the depth information, remove the background interference. And then using CamShift tracking algorithm only to track the removed background prospect moving object. This method is also ausing Kalman filter to predict the position and velocity of moving targets, using different scale factors make the linear combination of CamShift algorithm tracking results and the Kalman filter prediction results. The algorithm fully uses the space target location and velocity information, can overcome the change of ambient light and the similar colors of moving objects occlusion problem, improve the tracking reliability.