Research On Pedestrian Detection And Tracking Method Based On Deep Learning

With the dramatic increase of library personnel during peak periods,obtaining information on foot traffic and detection is crucial for management statistics.Traditional methods of foot traffic statistics rely on manual counting or rule-based algorithms,which can suffer from large errors and low efficiency.On the other hand,deep learning acquires complex feature representations from vast amounts of data,enabling the resolution of detection and tracking issues in various complex scenarios,such as libraries with diverse lighting conditions,occlusions,and pose changes.This approach offers an efficient and convenient method for monitoring and managing the flow of people in intelligent environments.Facing the dense human flow and complex background of libraries,how to efficiently and quickly detect and track pedestrians is the key research direction of intelligent visual recognition at present.Therefore,a study on dense pedestrian detection and tracking methods is conducted in this paper,with the following main research contents:(1)A lightweight DeepSort-based pedestrian flow counting system has been proposed to address the problems of missed detections and slow tracking speed in library scenarios.Given the high density of pedestrian targets and self-occlusions among pedestrians in a library setting,target loss and high algorithmic complexity often occur during pedestrian detection and tracking,leading to slow performance on small devices.To address these issues,the CIoU loss function and DIoU-NMS algorithm were employed to improve the detection algorithm.Additionally,a lightweight tracking algorithm was designed by combining the ShuffleNetV2 algorithm with the DeepSort network to reduce the model’s parameter size while maintaining good accuracy.Experimental results show that the improved algorithm can efficiently track pedestrian targets,with a model size reduced to only 5% of the original model,improving detection performance for occluded pedestrians and accurately counting pedestrian flow and walking speed.(2)Visual information can be greatly disrupted in pedestrian detection and recognition due to factors such as the shooting angle and lighting of surveillance cameras.To address the issues of low recognition rates for small pedestrian targets and poor detection accuracy for distant targets in monitored visual scenes,a multi-scale pedestrian object detection algorithm has been proposed.In the improved network model,effective feature fusion structures were designed to enhance the model’s perception of deep features,Res2 Net reconstruction algorithms were introduced to strengthen the utilization of fine-grained feature information,and spatial pyramid attention pooling networks were added to enhance the model’s multi-level feature expression ability.The improved algorithm was trained and validated on the Crowdhuman dataset,and the results show that the optimized algorithm achieves good accuracy and real-time performance,making it suitable for high-density pedestrian detection tasks.(3)To carry out intelligent pedestrian flow management in libraries more efficiently,a pedestrian identification and tracking system based on library scenes are constructed by combining improved pedestrian detection and tracking algorithms and using pedestrian reidentification technology and face recognition methods.From the experiments,it can be seen that the system designed in this paper has a good effect on the statistics of pedestrian flow in actual scenes such as libraries,and the accuracy rate of pedestrian identification is high,which has important reference significance for the intelligent analysis of pedestrian flow in modern libraries.