Research On Single Object Tracking Technology Based On Adaptive Block And Online Discriminative Classifier

The single object tracking based on video is one of the hot topics in computer vision, which is widely used in the fields of military and civilian. There are some factors leading to tracking drift, such as non-rigid object, random movement. And tracking is still a challenge in the presence of illu mination change, rotation, deformation, occlusion and so on. In order to achieve robust object tracking, the mature and stable key technology and methods are needed. The tracking algorithm based on the online discriminative classifier obtains better tracki ng performance than the previous algorithms, since tracking problems can be regarded as classification problem and updating classifier in real time makes a contribution to robust tracking. Dividing the object into several blocks is a common method to solve the problem of tracking under occlusion, which cannot adapt to the various objects and the variations in appearance. Therefore, research on how to segment object and apply online discriminative classifier to tracking algorithms have significant theory value and application value.Aiming at the problem of object tracking under complex scenes, the object model based on block and combing it with the online discriminative classifier to tracking are researched intensivly. First, an adaptive block algorithm is proposed. The DBSCAN clustering algorithm based on su per-pixel is used to improve the algorithm. On this basis, online discriminant classifier is used to tracking, which uses fused feature extracted from the object blocks to training local classifiers, and establishes the confidence map by discriminant results of local classifiers; Then, confidences of samples can be obtained by combining motion model and local classifier confidences, and the sample with the maximum posteriori probability is the tracking results; Meanwhile, the classifier updating strategy based on block to judge occlusion is proposed to ensure the reliability of the sample s. Experimental results show that the proposed method can better solve the problem which tracking precision is low caused by object partial occlusion. In order to further improve the tracking accuracy, the tracking algorithm based on global classifier and local classifier is proposed to solve occlusion. The global classifier and local classifier are learned by global features and local features respectively. In addition, the updating strategy is used to guarantee the reliability of classifiers. Experimental results show that the proposed algorithm obtains better tracking precision under complex conditions, such as object appearance variation, background interference and so on.