Improvement Of Target Detection And Target Tracking Algorithm Based On Computer Vision

Targets detection and tracking based on computer vision are important directions in the field of computer vision. The research focus on analyzing motion region, detecting and extracting foreground targets, tracking the objects, obtaining the character parameters of targets, thus getting the motion trail, realizing tracking. Currently, targets occlusion, targets' size changed and the color of background was similar to object had distraction, these problems has a great influence on the effect of objects detection and tracking. This paper analyzed the traditional objects detecting and tracking algorithm and put forward to improve the algorithm to solve above problems, then the testing result proved that the improved algorithm better than the traditional algorithm. The contents as below:(1) For targets detection, introduced three common detection methods in static scene,compared their merits and faults and then came up with fusion algorithm. The disadvantage of Gaussian Mixture-based Background Model is that the learning rate is a fixed value, so bring forward to update the value, improve the algorithm's adaptability to environment. GMM could extract the integrity object, but it's sensitive to the environment and noise; Symmetrical differencing method's calculation is simple, but this method cannot extract a full target. According to the traits of the two methods, this paper put forward to combine with the two algorithms, the improved method not only eliminated the cavitation that generated by symmetrical differencing method, but also suppressed the influence of noise to GMM. In addition, edge-detection algorithm extracted border has lots of information, and it can make good use of the object's spatial information. While the improved method just used the temporal information, so put forward to combine the edge-detection algorithm with the improved method. This method had a good use of spatiotemporal information. The simulation results proved that the fusion algorithm had a better detection effect.(2) For targets tracking, the Meanshift algorithm cannot self-adaption adjust the tracking window according to the size of target. Then the paper focus on introducing Camshift algorithm,it's based on Meanshift algorithm. The algorithm isn't sensitive to the shape and size change of object. But the algorithm is easily to lose the target, and it's very sensitive to similar color disturbance. So we put forward to improve the traditional algorithm to overcome the above defects.Initialize the detected foreground image as the tracking window, optimize the algorithm, so that the algorithm could select the target automatically. Then combined the improved algorithm with Kalman filter, the Kalman was used to predict the location of object in next frame, and this method resolved the problem that when the target was sheltered and completed tracking. The paper tested the algorithm to track the people and the car, proved that the improved method has abetter tracking effect.