Object Tracking Based On Motion State And Trajectory Prediction

Video object tracking is one of the key technologies in the field of computer vision,and has a wide range of applications in real-life such as video surveillance,autonomous driving and Unmanned Aerial Vehicles(UAV)scouting.In recent years,with the rapid progress of deep learning technology,object tracking research based on deep learning has also made great development,and scholars at home and abroad have proposed many excellent approaches to the problem of achieving accurate and robust tracking.In practice,however,there are still many problems with these methods,for example,tracker performance may decrease in complex environments such as illumination changes,scale changes,occlusions,deformations,and fast motion.How to achieve both high accuracy and robust object tracking is still of great research significance.This paper analyses and discusses the development of existing tracking algorithms.Based on this,the paper focuses on the motion state of the target in the video and the continuity of the target trajectory.The main research in this paper consists of the following two points:1)We proposed an object tracking method based on adaptive bounding boxes strategy and motion state redetection.The adaptive bounding boxes takes two prediction boxes with different branches as candidates,estimates the reliability of the candidate boxes by calculating the conditional probability of the occurrence of different candidates,and selects the more reliable candidate boxes as the output to solve the angular error problem caused by the inaccuracy of the foreground segmentation map.The motion state redetection module detects the reliability of the current frame tracking by comparing the current motion state with the historical motion of the target,based on the idea that "the same target should maintain similar motion states in adjacent time intervals".When a tracking failure is detected,the tracker will retrack using an updated template containing the feature of high quality frames in the history sequence.The method is validated on the VOT2016,VOT2018 and VOT2019 datasets,and the results show that the method achieves significant improvements in accuracy and EAO compared to the baseline method Siam Mask,and the EAO of this method are also at the cutting-edge compared to other methods.2)We proposed a redetection object tracking method based on target trajectory prediction and template fusion.The method uses trajectory prediction to verify the robustness of the tracker in terms of spatial continuity.When the tracker’s tracking trajectory deviates significantly from the predicted trajectory,the tracker considers that the current tracking may have lost the target.When a target loss is detected,the tracker uses a new search region cropped with the center of the predicted position and retraces using a fusion template incorporating the features of the nearest frame.To avoid tracking failure caused by prediction errors of the trajectory predictor,the tracker also compares the candidate tracking frames from both tracks after redetection and selects the more reliable of the candidate as the final output.The method is validated on three datasets,OTB2015,UAV20 L and VOT2019,and the results show that the method achieves a significant performance improvement in robustness compared to the baseline method Trans T,while the robustness and accuracy of the method is equally superior compared to other frontier methods.