Research On The Key Techniques For Low Level Light Video Target Tracking

Video target tracking is a key problem in computer vision with a wide range of applications in visual surveillance, military guidance, visual navigation of robots, human-computer interaction, and medical diagnose, etc. The goal of video target tracking is to simulate motion sensing capabilities of human vision in order to enable machine to identify the moving objects in image sequences, providing important data for video analysis and video understanding. Due to the complex background image and the variety of target movement, it is very difficult to develop a robust tracking algorithm. Especially in the low level light conditions, with heavy image noise and low image resolution, we almost cannot recognize the target from the background. Under this condition, it becomes more difficult to achieve stable,reliable, and robust tracking result. In this paper, we have studied the key technologies of video target tracking in the low level light condition. The main work includes the detail descriptions on the following:1. Describe the characteristics of low level light images, the basic image enhancement algorithm, the basic motion detection algorithm and the Kalman filter theory. 2. Propose a low level light image enhancement algorithm based on spatial-temporalthree-dimensional histogram. The algorithm constructs a spatial-temporal three-dimen- sional histogram by jointing the pixel and its spatial/temporal adjacent pixels. Then segment the image into two different regions with Otsu, Fisher criterion function and maximum information entropy algorithms. Do histogram equalization on the two regions separately. Finally, synthesize the two regions to obtain the complete processed image. The experimental results demonstrate that: the standard deviation, entropy, contrast promotion index increase at least 1.9%, 2.2% and 2.0% respectively, when the low level image enhancement algorithm based on three-dimensional histogram is introduced.3. Propose a mean shift tracking algorithm based on color and edge features. On basis of the combination of the advantage of color feature which is not sensitive to rotation and deformation of the target and the advantage of edge feature which is not sensitive to illumination changes and color changes, establish the color histogram and edge orientation histogram for the target, and take them as the target model of mean shift algorithm. Meanwhile, to further improve the search efficiency, Kalman filter is introduced to predict the target. Experimental results show that the new method is more accurate for tracking, and it works well when illumination changes or target deformation occurs, and the robustness is improved as well.4. Propose a tracking algorithm based on SURF features of target and local background. Models are established by extracting the SURF features of targets and local background simultaneously, as solves the problem that the target and background information overlapping could lead to tracking failure. And update the models to meet the apparent change of the target in the movement. The experimental results show that: compared with the tracking algorithm based on SURF features and the tracking algorithm based on SURF and color two features, when similar background interference and partial occlusion occurred, the average tracking error of this algorithm has reduced 52.1% and 34.1%, respectively. When similar background interference and horizontal rotation occurred, the average tracking error of this algorithm has reduced 68.3% and 23.4%, respectively.