Research On Pedestrian Detection And Tracking Algorithm Based On Surveillance Video

With the rapid development of deep learning,the performance of pedestrian detection and tracking models has been greatly improved.However,in complex environments,surveillance videos are prone to small object sizes and obscured pedestrians,which bring great challenges to the current pedestrian detection and tracking algorithms.Due to the less information of small objects is easy to miss detection,which leads to the decrease of detection accuracy and thus the tracking effect is poor.Therefore,there is a need to further improve the tracking performance based on the improved detection effect of small objects.To address this problem,this thesis investigates pedestrian detection and tracking algorithms in surveillance video,and the work revolves around both pedestrian detection algorithms and pedestrian tracking algorithms,as follows.(1)To address the difficult problem of small object detection,a pedestrian detection model combining cooperative attention mechanism and P2-Bi FPN(P2 Bidirectional Feature Pyramid Network)structure is constructed with YOLOv5 l as the benchmark model.First,the cooperative attention mechanism is added to the residual unit of the backbone network to improve the accuracy of small object localization,and the symmetric convolutional kernel 3×3 of the residual unit is replaced by parallel asymmetric convolutional kernels 3×1 and 1×3,thus reducing the number of model parameters.Then,the P2-Bi FPN feature fusion network is built to introduce the small object information from the shallow layer of P2 to the high level feature map through the feature enhancement module,thus enriching the small object information and further improving the detection efficiency of small objects.Finally,the algorithm is trained and tested on the Wider Person pedestrian detection dataset.The experimental results show that the m AP of the pedestrian detection model constructed in this thesis is improved by1.7%,and the amount of model parameters is reduced by 5.66 M compared with the benchmark model.(2)To address the occlusion problem between pedestrians,an enhanced data correlation anti occlusion pedestrian detection tracking method is designed based on the pedestrian detection algorithm.The pedestrian detection model combining cooperative attention mechanism and P2-Bi FPN structure is used as a detector to do anti-occlusion treatment to the original Deep SORT tracking algorithm.First,the Ghost Net-cot reidentification network built in this thesis is introduced into this algorithm model,which combines the lightweight Ghoset Net model and the Co T block network that introduces contextual information,so as to extract a more discriminative feature map and give higher attention to the pedestrian part.Then,the data association matching method is optimized to combine spatial and appearance information to match the detection frame and confirmation state trajectory of the occluded object again,thus improving the performance of pedestrian tracking in the occluded scene.Finally,the algorithm is trained and tested on the MOT16 tracking dataset.Analyzing the experimental results,it can be seen that compared with the original Deep SORT algorithm,the anti-occlusion Deep SORT tracking algorithm studied in this thesis has improved HOTA by 3.61%,reduced IDSW by 33%,and has better tracking effect.Compared with the benchmark model YOLOv5+Deep SORT algorithm,the HOTA of the surveillance video-based pedestrian detection tracking model constructed in this thesis is improved by 9.03%,indicating better tracking effect on small objects and obscured pedestrians.