Research And Application Of Long-term Object Tracking Algorithm Based On Ground Scene

Object tracking is one of the research hotspot in the field of computer vision,which aims at locating and tracking the object in a continuous image sequences.Long-term tracking is an important branch in the field of object tracking.Considering tracking deviation,it adds re-detection module.It has a wide application prospect in driverless vehicles,weapon guidance,UAV navigation,and video supervision.Compared with the sky scene and the ocean scene,the ground scene has more challenges,and in the process of tracking,it is easy to appear complex situations such as similar object clutters,scale variation,occlusion and out-of-view,which often leads to the failure of tracking.Due to its strong feature representation ability,deep learning is a hot topic in the field of object tracking,its representative is SiamFC algorithm.This thesis also carries out research based on deep research of this algorithm.This thesis analyzes the problems existing in the long-term tracking under the ground scene,proposes a series of improvement methods,and conducts the research and application of the algorithm.The main contributions of this thesis are:(1)To address the shortcomings of existing deep learning tracking algorithms under the interference of similar object clutters,based on SiamFC algorithm and combined with transfer learning,a object tracking algorithm based on instance transfer learning is proposed(TLSiamFC).In the first frame of the image sequence,based on the offline pretrained target recognition model,we implement instance transfer learning,extract the object instance discriminant features and scale-sensitive features,and then build the tracking feature extraction network for the object,combined with the Siamese Network,perform tracking task for subsequent frames.The experiments show that the improved algorithm improves the success rate and accuracy by 5.2% and 5.7% respectively,and can maintain the speed of 66.8 frames per second on the CPU,which effectively improves the tracking performance in the ground scene.(2)For the long-term tracking scenario,TLSiamFC algorithm is further improved.After analyzing the situation of tracking deviation caused by object occlusion,a longterm object tracking algorithm based on re-detection is proposed(LTSiamRPN).In the redetection mechanism,the peak detection of response score graph is added to judgment the tracking result of each frame.When tracking failure is found,the Region Proposal Network(RPN)is added to locate the candidate target accurately and recover the tracking object.At the same time,the object template update strategy is improved by combining the interval update and weighted update.The experiments show that the improved algorithm improves the success rate and accuracy by 2.8% and 5.4% respectively,and can maintain the speed of 23.5 frames per second on the CPU,which effectively improves the long-term tracking performance of the algorithm.(3)We implement the LTSiamRPN algorithm in the DSP embedded platform with DM6437 ZWT as the core.By processing the data set,lightening the network model and accelerating the algorithm on the platform,the algorithm is successfully applied to the mobile tracking system,and can maintain the real-time capability of 26.4 frames per second.Through scenario test,we verify that the algorithm has good practical value.