Research On Visual Object Detecting And Tracking Under Complex Environment

Visual object detection and tracking is one of the key technologies in computer vision,and is widely used in civil and military fields, including intelligent video surveillance,intelligent human computer interaction, video compression, automatic pilot, precision guidedsystem and so on, so the research of visual object detection and tracking has very importanttheoretical significance and application value. In recent years, with the development of rapidincreased computer processing power and image and video analysis techniques, researchers invisual object detection and tracking field proposed a lot of good algorithms, a variety ofobject detection and tracking application oriented techniques have emerged.For visual object detection, this thesis focuses on the object shape feature, and proposeda method based on kAS features. Polygon line segment approximation of the contour curve inthe preprocessing stage makes kAS characteristic descriptor more accurate; Block weightedkAS histogram in feature extraction gives full consideration to feature distribution, andimprove the accuracy of object detection; Block based sliding window in detection processmakes the calculation faster. Experimental results show that this method can get higherrecognition rate, with the property of translation and scaling invariance.For visual object tracking, through in depth understanding of the framework of MeanShift algorithm, we realized that Mean Shift is apt to make errors when tracking around thesimilar area, or lose the target if the target is occluded, thus this paper proposed an algorithmbased on RVM Mean Shift, moving regions are firstly extracted, track in the movementregions can effectively limits the Taylor expansion of the error caused by the Mean Shiftalgorithm; local information used in the establishment of target/candidate model makes thetracking more robust, while reducing the number of iterations of Mean Shift. Experimentalresults show that the proposed algorithm can track object successfully and have better robustfor occlusion.Finally, we make full use of the object detection and tracking with their respectiveadvantages, and then proposed a tracking algorithm based on Mean Shift and object detection,using Mean Shift algorithm as a tracker, random fern together with the nearest neighborclassifier as a detector, the final result is given by the fusion algorithm. The random fern classifier using the binary image features can be efficient to filter most of the background area,nearest neighbor classifier is used in order to further improve the classification accuracy.Experimental results show that the algorithm meets the real time tracking requirements, andthe use of detector makes the tracking effective in dealing with "fade","jump cut" as well asocclusion.