Research On Methods Of Object Detection And Tracking Based On Video Sequence Images

Target detection and tracking is a core issue in research field of computer vision and pattern recognition. With the target detection and tracking been applied widely in the corresponding field, some new and higher requirements have been put forward, which prompted the target detection and target tracking of in-depth study. In the process of research, people take more attention to the accuracy and stability of the algorithm in practical application, and the abilities to adapt to the complex environme nt. This thsis mainly studies the basic method of moving target detection, various features model of target and multi feature fusion,the basic theory of particle filter algorithm and Mean Shift method and the application process in target tracking, this thsis put forward the improvement method also.At target detection part, this thsis mainly on the inter frame difference method, and using the technology of digital image processing to detect moving target. O n detection of moving target in dynamic background, the introduction of corner detection and matching method, the image affine transformation, which compensates the effect of dynamic background and detected the moving target accurately in the experiment. To distinguish between pedestrians and vehicles in target detection, this thsis puts forward a method of using two phase proportion. O ne ratio is the target rectangle width to height. The other ratio is the longest line in the connected region which near the center of the target to the corresponding rectangle frame length, the effect is good actual testing.In target tracking, the target model has an important influence on the tracking results. This thsis introduces the extraction and modeling method of color histogram, LBP texture feature and edge directio n features, and puts forward the method of fusion of those features. Feature model fused background information, which is able to inhibited background effect for target tracking. Because the texture feature can cope with illumination variation, enhancing the target description ability and feature fusion target color histogram feature after.On the particle filter tracking process, particle degeneracy leads the search range become smaller, so as to fall into the local optimal problem. This thsis introduces t he K-means clustering on particle swarm optimization, so as to ensure the diversity of particle information, and find out the global optimal. In the process of tracking, the Bhattacharyya coefficient to target larger performance poor, this thsis improves the method of coefficient expression to make the tracking window can be adaptive to changes in object size. Experiments show that the mothed improve the tracking stability, can effectively deal with the size change and occlusion problem.Mean Shift is a non-parametric estimation method based on density gradient, through the method of iterative optimization gradually converge to a local extreme point. When the target moving fast, target occlusion occurred, and so on, the tracking effect is poor. Tracking fast moving target, the method did not have a predicted scope of the target, the calculation amount is relatively large, and even into the local optimal result in failure. This paper introduces Kalman filtering method which has the ability of state forecasting, Mean Shift method will start from the prediction region to find the target, reducing the number of iterations, while improving the stability. In response to the target occlusion, first through the coefficient estimate the occlusion problem takes place, generating method Kalman filter uses only the occlusion of state prediction so as to ensure the target is not lost. This thsis establishes a system of multiple target tracking method, this method matched the detected targets with the tracked target and processed object occlusion, loss and other issues, and final bring about stable tracking of multiple objects.