Research On Pedestrian Detection And Model Compression Based On Convolutional Neural Network

Pedestrian detection can be combined with pedestrian re-recognition,pedestrian tracking,which is widely used in public security,intelligent driving and other fields.However,the existing pedestrian detection network models mainly rely on general object detection models,and achieved great results.But applying direct the general target model detecting pedestrians,it causes missed detection and accuracy declining when the targets are occlusion.In this paper,based on the relevant theoretical knowledge of the convolution neural network and the general detection model,constructing a detection model suitable for partial occlusion of pedestrians,and proposing a method of pedestrian occlusion detection based on an improved YOLOv3 network structure,which can improve the efficiency of pedestrian detection.What's more,in order to apply in terminal equipments,we compress and optimize to facilitate pedestrian detection model.The paper's main tasks are as follows:(1)On the basis of comparing the existing network structure and network model,the representative features can be extracted by the convolution neural network.The deep learning method is used for pedestrian detection to extract more robustness so that it can apply efficient the next stage of pedestrian occlusion detection.(2)Aiming at the problem that the background is complicated and pedestrian occlusion in the real-time scene,pedestrian detection quality is low.This paper improves YOLO network based on the existing convolution network model,and proposes an partial occlusion detection algorithm based the network structures.By enhancing the shallow feature output of the network to improve local feature detection,and using PRe LU to replace Re LU,Finally introducing the GIo U loss function to solve the problem,which traditional Io U can't accurately feedback the overlap when the two targets are occluded and the intersections point are the same.(3)For some existing pedestrian detection networks it is difficult to deploy on the mobile terminal.By analyzing the existing model compression methods,the knowledge distillation model compression algorithm is proposed.According to use the improved YOLOv3 model as the teacher network T,after achieving a good training effect.The student network S learns knowledge from the teacher network T.And optimizes it based on the loss function.So the student network S is finally used for mobile terminal deployment.Experimental results show that the proposed detection algorithm can not only achieve real-time pedestrian detection,but also obtain better detection accuracy,which compared with other method.For the compression method based on the knowledge distillation model,Experimental results show that this method not only guarantees the original detection accuracy,but also can effectively reduce the parameter redundancy,which is beneficial to the model deployment.