Implementation And Application Of Multi-Target Recognition Tracking And Counting Based On Deep Learning And Hardware Acceleration

With the concept of a smart city,deep learning,especially target recognition and multiobjective tracking,has become its important research direction to optimize urban commuting problems through reasonable control of urban traffic and human flow.Real-time multi-objective tracking counting can well count the flow of people and vehicles,provide the necessary safety warning as well as diversion work for urban management,and provide strong data support for future urban road traffic planning,which has very broad application prospects.In this paper,we propose a multi-target tracking and counting method based on deep learning,with hardware acceleration to achieve real-time counting and design a visual application interface.The main work includes the following aspects:(1)The study is based on YOLOv5 version 6.1 target recognition algorithm,combined with the multi-target tracking algorithm Deep SORT to achieve multi-target recognition tracking.In order to improve the running performance,the target detection algorithm YOLOv5 s is optimized and improved,including adding the CBAM attention module,replacing PAN with Bi FPN structure,and changing the loss function CIo U to a more advanced SIo U to improve the feature extraction capability of the recognized targets.Finally,the Inception＿Conv is used to reduce the model parameters The improved YOLOv5s＿CBBi+IC +SIOU target recognition algorithm model is obtained,and the accuracy is improved by 1.5% compared with the original YOLOv5 s algorithm during the test.(2)The target tracking algorithm based on the improved YOLOv5 s and Deep SORT.In order to achieve the real-time requirement of tracking,this paper performs TensorRT engine acceleration on the improved YOLOv5 s algorithm,so that it calls the hardware Tensor Core unit of the graphics card for inference acceleration.Three kinds of data sets,namely Person class,Vehicle class,and Mix class,are collected and trained to obtain three recognition models.By selecting a single recognition model for the tracking count category,the recognition speed and accuracy can be further improved,while the Mix model provides both Person class and Vehicle class tracking recognition counts.The study shows the hardware-accelerated model has an average 63% improvement in recognition frame rate compared with the original model,while there is almost no significant impact on recognition accuracy.The accuracy of the three models reaches 96% on average,and the running frame rate is about 25 FPS,which can well satisfy the requirements of real-time multi-target tracking and counting.(3)A visual application interface is designed and implemented using TKinter,which includes the selection of three recognition models for deployment,corresponding to different recognition targets.Choosing to import local videos for counting statistics or calling hardware cameras for real-time counting statistics,and providing recognition result video saving function,the application can provide intuitive operation for users and guarantee the grounded application of the algorithm.