Research On Pedestrian Detection Based On Convolutional Neural Network

Pedestrian detection is an important task in the field of computer vision,and it has great research value and application prospect in the fields of vehicle driving assistance,video surveillance,intelligent transportation and intelligent robot.In recent years,pedestrian detection algorithms based on convolutional neural networks have developed rapidly,and the performance of pedestrian detection models has been continuously improved.However,high-precision pedestrian detection models often have some disadvantages such as large model volume,high computational cost and poor real-time speed,and it is difficult to adapt to the requirements of various application scenarios of pedestrian detection.The aim of this study is to improve the detection accuracy and speed of pedestrian detection algorithm,and then reduce the deployment cost of the model by combining model compression technology.This article improves the YOLOV4 algorithm and proposes a compression algorithm suitable for pedestrian detection model.The main work of this thesis includes:(1)In view of the mismatched parameters between YOLOV4 general target detection algorithm and pedestrian detection task,k-means++algorithm was used in this thesis to perform prior-box clustering on INRIA data set and Caltech data set,and the average intersection and combination of anchor boxes obtained by clustering was significantly improved compared with the k-means algorithm.In addition,Focal Loss function was used to replace cross-entropy loss function in the loss function design in this thesis to improve the imbalance of positive and negative samples of pedestrian targets.(2)To solve the problem of duplicate gradient information and model parameter redundancy,this thesis added cross stage partial design to YOLOv4 multi-scale fusion module.On the basis of not reducing the detection accuracy,the floating-point calculation amount of the detection model decreases by 15.6%and the model volume decreases by 17.9%,which effectively reduces the deployment cost of the model.(3)This thesis proposed a compression algorithm based on YOLOV4 pedestrian detection model.Through sparsity training,model channel pruning,fine-tuning training with knowledge distillation and other steps,large model compression is achieved at the cost of a small loss of precision.When the pruning rates are 0.1,0.3,0.5 and 0.8,the number of parameters in the compressed lightweight pedestrian detection model decreased by 8.6%,26.3%,44.0%and 63.4%,respectively,compared with the original model.