Multi-instance Learning Of Object Tracking Algorithm Based On SURF Feature

Video target tracking is an important research direction of machine vision and artificial intelligence,and has a strong practical value and important significance in theory research.So many scholars and experts are studying how to better solve the influence of target tracking of the lighting change,shape change,appearance change and occlusion.In order to improve the robustness of the algorithm,realizing tracking methods and ideas have been followed up.In this paper a target tracking algorithm based on the SURF and MIL is proposed.Firstly,we extract the SURF features of the target of interest and its surrounding image;secondly,SURF descriptor will be introduced to the MIL as the examples in positive and negative bag;thirdly,cluster all the extracted SURF features,and establish a visual vocabulary;fourthly,calculate the importance of the visual words in bag to establish a "word document" matrix,and get the latent semantic features of the bag;finally,train support vector machine(SVM)with the latent semantic features of bag,so that MIL problems can be handled in accordance with the supervised learning problem.We could determine where the target of interest among all observations is,and track the object efficiently.The experimental results show the robustness of the proposed algorithm under the situation of the variation of the scale,gesture and appearance,as well as short-term partial occlusion.In conclusion,we combine SURF and MIL for the target tracking.The proposed method can effectively realize target tracking.