Research On The Related Technology Of Target Tracking Algorithm

As a research hotspot in computer vision field,target tracking has widely prospect in application.Although the target tracking technology has made great progress in the past decade,designing a long-term target tracking algorithm that can deal with various complex situations such as illumination variation,geometric deformation,fast moving,occlusion,background clutter is still very tough.Recent years,the tracking-by-detection method is very popular due to its high efficiency and good performance.Tracking-Learning-Detection(TLD)target tracking algorithm as a typical representative of the tracking-by-detection method,decomposes the long-term target tracking task in three sub-parts,namely tracking,learning and detection.It has good real-time performance and high accuracy.Nevertheless,the TLD tracking algorithm has poor tracking ability and cannot track in cases including out-of-plane rotation,fast moving and non-rigid deformation.So,in the paper,several improvements proposed for the shortcomings of the TLD target tracking algorithm.Firstly,a classifier based on image perceptual hash is proposed,and candidate images are classified by their hash value.The image perceptual hash classifier serves as the third level of the cascade classifier of the tracking-learning-detection algorithm to overcome the shortcomings of the nearest neighbor classifier and enhance the detection performance.The experimental results show that the proposed algorithm has better tracking and learning accuracy than the tracking-learning-detection tracking algorithm,and the average frame rate is improved.Secondly,the median flow tracker easy to fail when the target's motion is large,so a tracking-learning-detection algorithm based on compressive sensing is proposed.The algorithm is used to improve the tracker,and compressive sensing is used for reducing the dimensionality of the extracted sample features to obtain the compressive features,then the compression feature is classified and the computational complexity is reduced;the compressive features retains most of the original information,so the tracker has better performance.The experimental results show that the tracking success rate and accuracy of the improved algorithm are improved compared with the tracking-learning-detection algorithm,and the character of real-time also improved greatly.Finally,owing to the tracking learning detection algorithm is easy to track failed in the case of rotating out of the plane,fast moving and non-rigid deformation,so a tracking-learning-detection algorithm based on kernelized correlation filters is proposed.The algorithm improves the tracker to boost its performance.In addition,exploiting incorporating color information in the tracking module to further enhance the tracking accuracy and robustness of the improved algorithm.The experimental results show that,In the complex tracking scene,the improved algorithm has a higher tracking success rate than the original algorithm.