Research On Object Tracking Based On Improved Camshift Algorithm In Dynamic Scenes

Visual tracking as an important research in computer vision field, widely used in the field of intelligent transportation, biomedicine, has important research value. There are so many excellent algorithms used by tracking including Camshift. Camshift algorithm which is tracked based on the color information, not only has strong robustness with translation, scale change and rotation invariance, but also plays an important role in the field of computer vision. But in dynamic scenes, complex background changes, occlusion, illumination changes or other factors will affect the accuracy of target tracking, even lead to the failure of tracking.In order to solve the problem of easy to be disturbed, firstly we improved the Camshift algorithm based on the common target detection and tracking algorithms in dynamic scenes, and secondly combine with Kalman filter to solve the occlusion problem. The main work in this paper are as follows:(1) Firstly, we analyze the target tracking problems and the interference factors in dynamic scene, next using high pass filtering method to filter the noise pixel in image sequence, and then improve the traditional setting pixel filtering threshold method, at last establish the contact between the window size and threshold to prevent the size of the tracking window mutation.(2) The traditional Camshift algorithm only extracts the H component while ignoring the impact of S saturation when convert the color space from RGB to HSV. In order to solve the problem above, this paper presents a two-dimensional histogram H-S based on H and S components to increase the utilization of color information. Considering the different contribution to color histogram in different pixels, give the pixels with different weight.(3) Occlusion is a common disturbance in visual tracking. Camshift algorithm is tracking object based on color information o, but the occlusion will lead to the failure of tracking. In order to solve this problem, we import Kalman filter to predict the location, but the traditional filter when in the face of large area occlusion will calculate the error result. Therefore this paper we import a large area of occlusion judgment, if not encountered the large occlusion predict the location by Kalman filter, else using the linear method to predict.At last, realize the algorithm proposed in this paper by computer vision library OpenCV and VC++6.0 in win 7 system, and then compared with traditional Camshift and improved Camshift combine with Kalman to verify the algorithm’s robustness and efficiency. Seen form the experiment results, the improved measures in the algorithm proposed in this paper, not only improves the robustness of the tracking system, but also improve the real-time property, as well suitable for target tracking in dynamic scene.