Research On Moving Object Tracking Based On Vision

Visual based object tracking has been applied very widely in the fields of video surveillance, image compression,3-D reconstruction, robot technology and so on. Thus, it is of great significance to study object tracking based on vision. The difficulties of object tracking lie in the sudden movements, the sudden change of the external performance form of the targets or the background, non-rigid structure of the targets, the block among goals, the block between the objects and the background, and cameras movement. To solve some of the above tracking problems, this dissertation focuses on how to improve the perfermance of object tracking. The correspondence theories and methods are deeply researched, and these fruits will provide a theoretical basis and technical support for the real application of object tracking. Four aspects of work are studied in this dissertation, as described as follows:(1) Since camera-motion changes background, method of camera-motion compensation is studied. Camera-motion compensation is a coordinate conversion procedure and it is inferred by corresponding salient features in the consecutive frame. The compensation can eliminate the background error greatly. An improved Gaussian model algorithm is proposed, which brings risk decision into the sudden judgment of goals of the prospect, and the performance of tracking is improved. Based on the view angle of the camera, a binocular vision localization method is proposed. To avoid the camera calibration, we analyse the two images from different locations using the camera view angles and calculate the actual coordinates in a three-dimensional space through the coordinates of target point pixel in the image.(2) When another object is in front of the object which we are tracking, or there are background disturbances, Mean Shift algorithm will slow down the tracking rate or lost the object. Weights are given to the background and to the object. The weight of background improves description of the object's characters. Different weights of different parts of the object increase Bhattacharyya value. After analyzing object template, the weights are brought to the mathematical expression of Mean Shift. The effect of tracking is better than that of the old one. (3) In order to improve the precision of tracking in real-time, the bandwidth of tracking should be changed according to the change of object's scale. Based on mean shift algorithm and the target center in every frame located by probability method, this thesis puts forward a self-adaptive algorithm for updating views. Then by combining the normalized moment of inertia (NMI) feature, the center of tracking object is located in real-time. The experiments show that this method can choose the proper size of tracking bandwidth when the object size changing and have strong anti-jamming.(4) To solve the problems of the traditional particle filter based on color informationunder the deformation of targets or sunshine, a particle filter object tracking algorithm based on multi-information fusion is proposed. The texture, color and movement charactors of object are fusioned. Experimental results show that the algorithm can accurately track moving targets on a strong anti-interference and robustness with a little increase of computation.