Detection And Tracking Of Moving Object In Complex Background

Moving targets detection and tracking is an important research aspect in the field of computer vision. It relates to several techniques, such as computer image processing, artificial intelligence and pattern recognition. The research of moving targets detection and tracking has many important potential applications, such as military affairs, motion analysis and so on.Based on the traditional process of target detection, setting up the model, the updating of model parameter and the sensitivity of the model to the scene etc, affect the speed of detection. More than, the shadow of the moving target affect the accuracy of the detection. Tracking targets in the complex background, the traditional algorithms are often tracking fail caused by the light changing, the high color similarity between the moving target and background. To solve these problems, this thesis studied some key problem of the targets detection and tracking algorithm for the complex sence, the main contents and contributions of this thesis are summarized as follows:(1) A Gaussian mixture model self-learning algorithm is proposed, the recursive equations of forgetting factor and learning rate factor are obtained based on EM algorithm. The updating parameters are more accurate and the convergence speed is much faster. The shadow is produced by the light, and detected in the process of detection, so we use the HSV color model to eliminate the shadow, and make the target detection accurate. Experimental results show that the traditional algorithm detects the moving target at the 40th frame, and the gaussian mixture model self-learning algorithm detects the moving target at the 22th frame. After eliminating the shadow, the prospect target is more accurate.(2) The interacting MCMC particle filter algorithm is proposed, which avoids sample impoverishment, at each time step the introduction of the interaction of particles reduces the correlation among one particle's history states in the method, and speeds the convergence rate. We use the results of detection, and choose two diagonal vertexes containing moving target boundary of region as tracking feature points, the algorithm is used to predict and track the position and velocity of the feature points, the thesis chooses the 3D space position and velocity as state variables. It avoids tracking fail caused by a nonlinear function. Experimental results show that the tracking fail doesn't happen in the complex environment, when the light changes, the moving target and background have high similarity. Also it can predict and track the target in 3D space trajectory. (3) A new algorithm combining the fuzzy Data Association and Particle Filter algorithm is proposed and used in the multi-target tracking, the algorithm joins the improved fuzzy functions after the resampling, membership degree is taken as the weight of the particle in filtering, which effectively avoids the interference of noise data. After the analysis of the algorithm, the targets are not tracking fail, when the moving trajectorys have overlapped. Experimental results show that the algorithm tracks the targets not only in the simple environment, but also in the complex environment, and enhances the robustness and accuracy of the targets tracking.