Pedestrian Detection Based On Convolutional Neural Network

With the development of intelligent driving,intelligent monitoring,and intelligent robots,pedestrian detection technology has attracted more and more attention.It is the cornerstone of research on target tracking,attitude analysis,behavior recognition,and pedestrian identification.The purpose of pedestrian detection is to detect pedestrians appearing in the image in real time,and give the specific location of pedestrians with rectangular boxes.Pedestrian detection technology faces enormous challenges due to the intensity of light,varying angles of shooting,complex pedestrian poses,and other objects in the image that resemble pedestrians.Therefore,pedestrian detection technology has always been a research hotspot and a difficult point in the field of computer vision.The main work of this thesis is to improve and perfect the pedestrian detection method based on the convolutional neural network.The main research contents are divided into the following two points:(1)The traditional feature extraction method is difficult for pedestrians in a highly variable environment to obtain feature information with high discrimination.At present,the popular convolutional neural network training based on back propagation algorithm is easy to fall into the dilemma of local minimum,which affects the classification performance.Aiming at the above problems,the pedestrian detection algorithm combining depth perception and kernel extreme learning machine(KELM)is proposed.First,on the CNN infrastructure,the front layer features and the deep layer features are combined in two stages to be the input of the subsequent layer,namely the DAG(Directed acyclic graph)network.This thesis uses the nuclear extreme learning machine to classify the depth perception characteristics of pedestrians in the DAG network,and uses K-fold cross-validation to select the most suitable parameter pairs.In the detection stage of the complete image with the background,the feature map learned by the DAG network is combined with the GBVS(Graph-Based Visual Saliency)saliency detection algorithm to quickly mark the area containing the pedestrian in the test image,and then use multi-scale sliding window to detect the precise location of pedestrians in the labeled area.The experimental results show the effectiveness of the proposed algorithm.(2)The popular detection method at present uses region proposal network(RPN)to generate candidate regions,and then classifies the regions,we find that there are still many wrong candidate regions in the results of existing RPN networks.Inspired by visual attention mechanism,this thesis proposes to use visual attention mechanism to guide RPN network to generate more accurate candidate regions containing potential targets.The visual attention map is generated by the visual attention network constructed by the attention module and the convolution LSTM in the residual attention network,and the attention map is normalized.After the treatment,a score map between 0 and 1 is obtained.The score map is then multiplied with the feature map of each feature channel generated by the underlying network to obtain a new feature map,which is sent to the RPN network to generate a more accurate candidate region containing the potential target.The features corresponding to the regional recommendations,confidence,and regional recommendations are then used to train the cascade enhanced forest classifier to obtain the final result.The algorithm in this treatise has achieved good results in the Caltech dataset and ETH dataset.