Adaptive Feature Fusion And Template Update Tracking Algorithm

Video target tracking is achieved by extracting the features of the target object,and then detecting the moving target in the sequence of video frames,thereby obtaining information such as the location and motion path of the target,and has been used in multiple fields like security monitoring,traffic monitoring,and defense reconnaissance.The target tracking technology continues to make breakthrough,developing from traditional algorithms such as Meanshift,Particle Filter and Kalman Filter,to the latest algorithms with theories of correlation filtering and deep learning.However,due to the complexity of the target tracking itself,the existing tracking algorithms are always unsatisfactory.Although the traditional tracking algorithm is small in computation and fast in speed,the accuracy of it is relatively low.While deep learning algorithm is good at accuracy and robustness,but because it requires large amount of deep learning,it is not always in real-time and also it demanding for hardware.Therefore,this paper focus on the algorithm balance between accuracy and speed in the process of tracking.Through analysis and research,the accuracy and robustness of the correlation filter tracking algorithm are much improved compared with the traditional algorithm,and the algorithm is faster.Therefore,this paper studies this algorithm.Through the analysis of the relevant filter tracking algorithm,this paper mainly discusses two aspects:feature selection and fusion,and template update.(1)The traditional single-target tracking algorithm generally uses only one feature to describe the target object.However,the single feature does not fully represent the moving target,and the target object is often interfered by the complex environment during the tracking process.Therefore,it is necessary to achieve tracking of the target by merging multiple features.Based on the above aialysis,this paper introduces the direction gradient histogram feature and the color histogram feature,trains different templates for different features,and then obtains their response graphs respectively.Because the fusion of fixed weights is less adaptable to complex environments,this paper adopts an adaptive feature fusion method.(2)Another problem in the target tracking process is target drift,and the most important reason is the error training and update operation of the feature model,which causes environmental errors in the subsequent tracking process,and finally causes tracking failure.Therefore,before the operation of updating the feature template,the confidence of the current frame matching result is first determined.When the confidence is greater than the set threshold,the result of the current frame is relatively reliable,and then the template is updated,otherwise the operation is discarded.