Research On Object Trackingalgorithm Based On Region Feature

Visual object tracking has been a hot research issue in image processing and artificialintelligence and playing an important role in intelligent surveillance, human-computerinteraction, visual navigation and precision guidance. Although numerous visual objecttracking algorithms have been proposed, there still exist a number of difficulties in realapplications, such as scaling, appearance variations, object occlusion and illuminationchanges of surroundings. To deal with the above mentioned problems of visual objecttracking in practical applications, several novel researches have been conducted onregion-based object tracking in this paper. Research and innovative achievements includefour aspects as follows:(1) To address the drawbacks of object appearance variations in infrared imagetracking, an object tracking algorithm is proposed based on adaptive multi-feature fusionand Mean Shift tracking model. Firstly, gray-level features and local mean contrast featuresare exploited to represent the object, and the uncertainty of features is introduced in featurefusion. Secondly, the algorithm utilizes the geometrical features and gray features as thedescriptors of object scale factors which are used to update the tracking window. Finally, aweighted template updating technique is adopted to update the object templates to improvethe robustness. Experimental results demonstrate that this proposed algorithm caneffectively adapt to the changes of image scenes and object appearances, and has goodperformance for infrared object tracking.(2) To settle the existing issue of object appearance variation in visible image tracking,a scale adaptive and multi-model compressive tracking approach is developed based oncompressive tracking model. An adaptive scale method is firstly proposed to solve theproblem of fixed scale. Then, multi-model fusion and classifier training are employedinstead of single model. A plenty of experiments are conducted to verify the feasibility andeffectiveness of the proposed method. Experimental results show that the improved CTalgorithm has good performance in object tracking with changes in scale and appearance.(3) In order to improve the tracking robustness with large-area occlusion, a newanti-occlusion object tracking algorithm based on Mean Shift and step-by-step localizationis proposed. The method can predict whether there exist occlusions or not. If the target isnot occluded, the original Mean Shift model is thus applied in tracking, and meanwhile an object template updating strategy based on selected components is used to reduce theinfluence of template drifting. When the object is occluded, candidate region modeled withNon-symmetric kernel is firstly utilized to reduce the occlusion influence of templatematching. Fragments-based model is then used for finding positions of the object andIntegral Histogram is employed for lowering the computational complexity. Moreover, animproved candidate model is adopted to compute the matching coefficients of object, whichavoids excessive object searching. Extensive experimental results show that the algorithmhas good tracking performance for large-area occlusion object.(4) To meet the requirements on operational capability, robustness, volume and powerconsumption for infrared object real-time tracking system, a high speed object trackingsystem based on SOPC is designed and realized. The system contains highly integratedFPGA hardware platform and software algorithms, which implements image acquisition,image non-uniformity correction, image filtering, feature extraction, object detection andtracking in real time. The provided experimental results demonstrate that the system canfully meet practical requirements and has good tracking performance.