Research On Problems Related To Adaptive Detection And Tracking Of Moving Objects

In recent years,aiming at the problems of traditional video processing,such as low intelligence,large processing error,inability to detect and track,etc,more and more video surveillance scenes use computer vision technology.The introduction of computer vision technology makes video processing more intelligent,efficient and user-friendly.Target detection and tracking is a research hotspot in the field of computer vision.It has important applications in human social production and activities.Its main task is to find and locate targets,and then track and analyze the targets.However,in practical application scenarios,there are many challenges such as target occlusion,scale variation,tracking drift,complex background and so on,which cause interference to detection and tracking.Aiming at the problems in actual application scenarios,this paper analyzes and researches from the two aspects of target detection and target tracking.According to the advantages and disadvantages of existing algorithms,an improved moving target detection and tracking algorithm based on You Only Look Once(YOLOv3)algorithm and Kernel Correlation Filter(KCF)algorithm is proposed,and discusses the specific application of lead fish detection tracking monitoring system.1.Aiming at the large amount of real-time calculations required by YOLOv3 in the process of target detection,ordinary computers can't handle a large number of real-time calculations,and the real-time performance of detection is poor,which can't meet the requirements of realtime detection of moving targets,propose the adaptive detection network structure M?YOLO for moving targets,adjust and reconstruct the network structure of the YOLOv3-Tiny algorithm for strengthen the detection ability of small targets,and at the same time merge the parameters of batch normalization(BN)layer into the convolution layer to reduce the calculation amount and speed up model prediction.The improved M?YOLO algorithm improves the accuracy rate by6.73 % compared to YOLOv3-Tiny,reduces the missed detection rate by 5.21 %,and reduces the false detection rate by 1.61 %.Compared with YOLOv3's 14 FPS to 33 FPS,the detection speed has been improved significantly.In summary,M?YOLO balances detection accuracy and real-time performance.2.Aiming at the problem that the KCF algorithm in the tracking process can't auto adjust the target scale and can't deal with the target occlusion,propose the adaptive tracking algorithm Moving?KCF(M?KCF)for moving targets based on the KCF algorithm.This algorithm uses a dual-scale estimation strategy based on scale pyramids to achieve target scale adaptation without significantly affecting the tracking speed.At the same time,peak to side lobe ratio(PSR)is also introduced to determine whether there is occlusion or tracking failure in the tracking process.When the PSR is less than a certain threshold,the target tracking is stopped,the target template is reselected,continue to track after the target is initialized,increase the algorithm of tracking accuracy and improve the tracking success rate.Compared with the KCF algorithm,the improved M?KCF algorithm improves the accuracy by 6.6 % and the success rate by 8.5 %under the OPE evaluation method.3.Aiming at the problems of poor robustness and tracking accuracy of traditional target tracking algorithm in complex application scenarios,as well as the poor real-time performance of detection algorithm based on deep learning framework,this paper proposes an improved adaptive detection and tracking algorithm M?YOLO?KCF(Moving?YOLO?KCF)for moving target based on M?YOLO and M?KCF,it not only solve the problem that the KCF algorithm needs to select the target manually when it is used for tracking?target loss caused by fast motion and drift caused by error accumulation,but also solve the real-time problem when yolov3 algorithm is used for tracking.Through the analysis of experimental results and performance in practical application scenarios,the proposed improved algorithm M?YOLO?KCF can accurately detect the target and track it in real time,and achieve the self-adaptation of target detection and tracking,which have an excellent performance in actual lead fish detection and tracking application scenarios.