Research On Pedestrian Detection Method In Dense Scene Based On Lightweight CNN

The crowd scene is the key focus area of access inspection,disease prevention,control and emergency evacuation.It is the place where identity verification,passenger flow statistics and directional track are needed.Pedestrian detection in dense scenes can help identity interconnection authentication,cross mirror tracking and the construction of counting system,facilitate the construction of smart city.The emergence of deep learning brings opportunities to the field of target detection and provides new ideas for pedestrian detection.However,high-precision deep learning models often bring high complexity and high computation,which hinders model deployment and embedded application.In view of those question mentioned before,pedestrians in dense scenes are taken as the research object,and reasonably designs the YOLOv4-tiny algorithm,including the optimization of parameters and structure.Firstly,format conversion is carried out for Pascal VOC,self-made data set and Wider Person data set.Reasonable parameters are selected through kmeans++ algorithm.When the K value is 6 and 9 respectively,the anchor frame of the pictures with input sizes of 320,416 and 608 is set.Secondly,the improved YOLO-L algorithm is obtained by modifying the network structure through three optimization techniques: improving the multi-scale prediction layer,improving the backbone network and adding attention mechanism.Then,training and ablation experiments are carried out for YOLO-L,and the effects of each optimization point are compared.The results show that when the input size of Pascal VOC data set is 320,the detection accuracy of YOLO-L is 4.68% higher than that of YOLOv4 tiny.The accuracy of the pictures with size of 416 is improved by 2.28%.The accuracy of pictures with size of 608 is improved by17.84%.Similarly,the verification is carried out on the self-made data set,and the detection accuracy is improved by 9.63%,13.44% and 20.85% respectively on the samples with size of 320,416 and 608;On the Wider Person dataset,it increased by 4.44%,5.41% and 4.24% respectively.In order to facilitate the embedded application,this paper simplifies the structure of YOLO-L and uses the model pruning method for lightweight design.The main methods are as follows: use data from Wider Person for basic training and sparse training,use the pruning strategy to simplify the structure,verify through layer pruning and channel pruning,and fine-tuning training is carried out.Finally,the lightweight model YOLO-L-lite is obtained.The experiment shows that the detection accuracy of YOLO-L-lite in the test set is 0.4657;The size of the model is 7.61 M,which is 38.7% less than the original YOLOv4-tiny;The model complexity is 3.620 BFLOPs.Finally,the embedded application of YOLO-L-lite model is completed by using Jetson TX2.The single picture detection speed of the model on the development board is 0.061435 s,which can achieve real-time detection and meet the embedded deployment requirements of YOLO-L-lite.To sum up,a new deep learning algorithm YOLO-L is designed through structural optimization,its lightweight algorithm YOLO-L-lite is obtained by model pruning,the embedded application is realized by using Jetson TX2 development board,and this method provides a deployment scheme for the practical application of pedestrian detection.