The Target Tracking Algorithm Research Based On Multiple Instance Learning And Template Matching

Target tracking technology is a popular research direction in the field of artificial intelligence,the purpose of tracking is to verify the position of the target in the video sequence. This paper analyzes many popular tracking algorithms, through the study of the experimental simulation and contrast analysis, made a more detailed analysis and research on the decision-making model tracking algorithm and matching tracking algorithm respectively. Sudden movement and the changes of light has a certain robustness in the decision model of target tracking algorithm, but if the target is covered half or all, it is easy to lose the target. In the matching tracking algorithm, it has a certain robustness in pose change, rotation, occlusion. But it assumes that the target is different from the background at the first, once appear the similar background to the target, it is easy to arouse failure, and it can not timely response to the sudden changes of illumination. The paper combining the advantages of the two types of tracking algorithm for target tracking research,based on the analysis of two kinds of algorithm.First, focuses on the target tracking algorithm based on template matching, included three basic steps: build the model, matching tracking and update the model. Because it is difficult to describe the target in single feature, and will lost the spatial information when using the histogram of characteristics, An effective object tracking method was proposed to fuse multiple features and spatial information in an object model. The gray value features, texture features and edge features of the target object were combined in the joint distribution field descriptors to construct the object model. in the new frame, using model matching method to search the most similarity candidate target, achieving the target coordinate. At the same time, considering in the target model, there are usually the same characteristics gathered and form characteristic of the concentration areas, these features base on the target model are important to distinguish between target and background.Therefore, when constructing the target model, the characteristics of the aggregate distribution layer multiplied by a weighted coefficient, highlight the relative characteristics, aims to improve the stability of target tracking.Combining multiple instance learning and mean shift tracking, need do some optimization on their respective algorithm. In order to improve the classification accuracy of instance learning tracking method, to get more accurate position. Put forward the idea that using the integrated study,random sampling, generate multiple classifiers, integration them to determine the location of the object. For Mean Shift tracking algorithm, through combining the color features and texture features to construct target model, taking advantage of the features of each to build a robust target model, improving the tracking precision.Aiming at how to combine decision algorithm and matching algorithm, by the improved tracking of the multiple instance learning, getting the target position in a new frame, then judge whether tracking failure occurs, if tracking failure, Then mean shift tracking was used to revise the position, the judgment standard uses the displacement of the target distance, because the offset between two frames will not be too big, if the offset is too big, judge the tracking failure, then call Mean Shift tracking algorithm to revise the target location.