Research On Moving Target Detection And Tracking Algorithm In Complex Condition

Target detection and tracking technology, as one of the important research topics of computer vision, which combines image processing, pattern recognition, artificial intelligence and so on, is widely used in military guidance, intelligent transportation, aerospace and other important areas. In recent years, a lot of efficient algorithms have been proposed, but in real-time scenario, the deformation, lighting changes, shadows interfere, occlusion and so on is still the critical factors of affecting the robustness. This paper do some research from three aspects:image pre-processing, target detection, target tracking, and the improved detection algorithm based on the Gaussian mixture model and improved mean shift algorithm are proposed.This paper introduces the pre-processing techniques and makes a detailed analysis of the advantages of these methods for improving image quality, which provide a theoretical basis and quality assurance for subsequent target detection and tracking.In the aspect of target detection, the paper introduces the common detection method, the common background modeling methods, the Gaussian mixture model, and puts forward the problems of the traditional Gaussian mixture model. Any solutions are proposed:the four neighborhood online update learning rate method is proposed to deal with illumination changes. In terms of the mutations, a parameter pt and a stable constant factor φ2are imported. In terms of shadow elimination, the maximum color difference method is used, by using chromaticity difference instead of gray difference. Experiments show that the improved method can obtain significant results.In the aspect of target tracking, the paper makes a detailed research of the mean shift algorithm, including the density estimation under multi-dimensional, mean shift vector and so on. Then, an improved mean shift algorithm is proposed, vertical projection method is used to get proper fragments in the algorithm, different kernel weighted histograms are built within the target area, make the similarity measurement between the target template and the candidate template, and then use the mean iterative to located target. During tracking, using the back projection operation of components to judge the target deformation and occlusion, and timely update the target template and sub-block weights. Experiments show that the algorithm has good robustness.