Pedestrian Detection Based On Shallow Feature Fusion Guidance Deep Network

In recent years,the development of video surveillance and intelligent transportation systems has become more and more open,providing a broader space for the advancement of society.Pedestrian detection technology is one of its most important contents,generally refers to the use of computer vision correlation algorithms to locate and detect targets based on relevant information acquired by images or video frames.Pedestrian detection is widely used in the fields of intelligent driving,intelligent transportation systems,robots and advanced human-computer interaction,and is closely related to people's lives.Especially in traffic accidents,most of the casualties are pedestrians on the road.How to improve the detection performance of pedestrians and ensure the safety of pedestrians has attracted the attention and research of more and more relevant experts.Pedestrian detection has a great development prospect,but the object pedestrian positioning,detection are more difficult.The changes in the shape,posture,appearance and clothing of the human,the diversity of the scene,etc.;the influence of external environmental factors such as weather changes,illumination,and pixels all cause a certain degree of trouble for pedestrian detection.With the deep learning method,many breakthrough results have beenachieved in pedestrian detection,which can improve the performance of pedestrian detection to some extent.Therefore,this paper mainly conducts research experiments on pedestrian detection based on deep learning.Therefore,this paper mainly studies the following aspects:(1)Network design for deep learning pedestrian detection.Firstly,based on the Faster R-CNN algorithm framework,in order to select a better network,research and analyze two kinds of convolutional neural networks VGGNet and ResNet popular in the object detection field using two different types of data sets;Secondly,the study compares the influence of different parameters on the network;finally,the pedestrian detection network structure is designed to provide the basis for the later research.(2)Pedestrian detection based on shallow feature fusion guidance deep network.Aiming at the problem that the current pedestrian detection algorithm misses detection when the objects is too small in a complex scene,a method combining shallow feature fusion and deep convolution network features is proposed.Firstly,in order to extract the high-quality features of pedestrians,research the hidden layer characteristics of deep convolutional neural networks.Extracting two common pedestrian traditional features from the first block of the convolutional neural network—the HOG feature,the improved LBP feature,and the deep convolutional network features.Second,with the help of the Region Propose Network(Region Propose Network),RPN)advantages,the design chooses the anchors suitable for the pedestrian data set;finally.Target candidateframe classification and prediction are performed using efficient pedestrian characteristics.It was found that on the Caltech pedestrian data set,it was possible to effectively detect smaller target pedestrians and close pedestrians in the distant scene in complex scenes.(3)VGGNet network pedestrian detection implanted in the BN layer.Aiming at the current situation of pedestrian detection of occlusion of pedestrians at the close-up,a method for improving the VGGNet network structure is proposed.In view of the fast convergence of the BN layer and the high efficiency of the learning model,the BN layer is added to each convolutional layer of the network.then,analyzing and predicting of pedestrian detection.It was found that on the Caltech pedestrian data set,this method can improve the detection performance to a certain extent and effectively detect the occluded pedestrians.