Research On Pedestrian Tracking Algorithm Based On Deep Learning

In recent years,AI and computer vision have become hot topics,attracting the close attention of a large number of experts and researchers,and triggering a boom of extensive and in-depth research on them.Pedestrian detection and tracking are the core technologies in the field of computer vision,such as unmanned driving,intelligent security and human-computer interaction.In video surveillance,pedestrians are the focus of attention,and target tracking algorithms emerge in an endless stream.Among them,pedestrian tracking algorithm occupies the mainstream of tracking and is a hot research topic today.As a result of the pedestrians is not a fixed posture target,tracking,inevitably occur between pedestrian block each other or other objects blocking,these problems to a large extent affected the pedestrian tracking technology applied in the real scene,has a very robust pedestrian tracking algorithm has been the difficulties of the study,according to the above problem,this thesis research content is as follows:Building a pedestrian detection network model YOLO-P based on YOLOv3 optimization.Based on the target detection algorithm YOLOv3,it was optimized,including Mosaic enhancement of the input data to expand the number of training data sets.In terms of anchor box setting,adaptive anchor box calculation is adopted to automatically calculate the appropriate size of anchor box.Adding CSP structure to the trunk feature extraction network can improve the learning ability of convolutional neural network,reduce the memory cost and save space.When calculating the overlap degree between the target box and the prediction box,IOU＿Loss is changed to GIOU＿Loss,and the loss function is modified to make it more conducive to the optimization of the network.The optimized algorithm is YOLO-P pedestrian detection algorithm.The pedestrian detection database and self-annotated image data were used to train the network model.After the completion of the training,the public data set and a total of 6000 images annotated by oneself were used for testing.During the testing,it was guaranteed that the tested pictures did not appear in the training set.Building an improved pedestrian tracking model Deep-Sort-Sup based on Deep-Sort.In terms of determining the location of the target pedestrian,a better-performing pedestrian detector YOLO-P is added to the Deep-Sort pedestrian tracking framework,and the position prediction information of the Kalman filter on the target pedestrian and the pedestrian detector YOLO-P are compared to the current frame The detection information of the target pedestrian can be fused to more accurately determine the location of the target pedestrian.Then use the Hungarian matching algorithm to associate the target pedestrian to determine whether the target pedestrian in two adjacent frames is the same tracking target.When encountering occlusion,use the SIFT feature point to match the target with only a small number of points.Whether it is the same target pedestrian,increase the robustness of the network when the target is blocked.This matching strategy can solve the occlusion problem to a certain extent.Compared with matching only by the Hungarian algorithm and adding the SIFT feature matching strategy,the accuracy of the target pedestrian matching successfully increased by about 6.2%.Through verification,compared with Deep-Sort,the tracking accuracy of Deep-Sort-Sup is improved by about 4.9%.