Research On Pedestrian Detection Algorithm Based On Convolutional Neural Network

In recent years,with the development of deep learning and convolutional neural network,pedestrian detection algorithms based on convolutional neural networks have also achieved excellent detection results,but these algorithms still have some unsatisfactory detection effects such as occluded pedestrians and small-scale pedestrians.Therefore,from the perspective of pedestrian occlusion and different pedestrian scales,this paper designs and proposes two pedestrian detection algorithms,and finally combines the detection results of the two algorithms to obtain more accurate pedestrian detection results.The main research contents are as follows:First of all,we add the convolution block attention module CBAM and the twoclassification module on the basis of the pedestrian detection algorithm CSP algorithm based on the anchorless frame to get the AS-CSP algorithm.We first deepen the backbone network of the CSP algorithm from Res Net-50 to Res Net-101,and use it as the feature extraction module of the new algorithm.Then,the convolution block attention module which can enhance the feature expression ability of pedestrians and small-scale pedestrians is used as the attention module and added to the feature extraction module.Later,after the detection head module of the original algorithm,a two-classification module based on score fusion mechanism is designed to improve the algorithm’s ability to detect occluded pedestrians by increasing the confidence score of occluded pedestrians and reducing the confidence score of false pedestrians.Finally,in order to verify the performance of the AS-CSP algorithm,a campus pedestrian data set was produced.This data set and the City Persons data set are used as a validation set to verify the effectiveness of the algorithm.The experimental results show that the AS-CSP algorithm has achieved better detection results than the original algorithm in different scenarios such as general pedestrians,small-scale pedestrians,and occluded pedestrians.Secondly,we add the multi feature pyramid network MFPN to the single-stage pedestrian detection algorithm ALFNet based on the anchor frame regression mechanism,and at the same time improve the loss function of the algorithm to obtain the MF-ALFNet algorithm.We first improve the Bi FPN feature pyramid network to obtain the multi feature-feature pyramid network MFPN,and add it as a feature fusion module to the feature extraction module of the ALFNet algorithm.The network can improve the algorithm’s ability to occlude pedestrians by fusing the feature information of pedestrians and occluded pedestrians.Then,in order to solve the problem of the difficulty of separating pedestrians in the detection of occlusion within the class,the original regression loss function is introduced and added with the repulsive regression function,so that the convolution prediction module of the algorithm can separate the pedestrians that are occluded from each other during pedestrian detection,thereby improving the detection effect of such pedestrians.We also use the City Persons dataset and self-built campus pedestrian dataset to verify the detection performance of the MF-ALFNet algorithm.The experimental results show that the detection performance of the MF-ALFNet algorithm is improved compared with the ALFNet algorithm,and the detection performance of occluded pedestrians is particularly improved.Finally,we use the weighted detection frame fusion method WBF to fuse the detection results of the AS-CSP algorithm and the MF-ALFNet algorithm,and use the evaluation indicators and visual detection results to evaluate the fusion experimental results.The experimental results show that the log-average missed detection rate of the fusion detection results on each subset of the City Persons data set is better than the two algorithms,and the visual detection results also show that the fusion detection results take into account the advantages of the two algorithms.In summary,this article analyzes the phenomenon of poor detection of occluded pedestrians and small-scale pedestrians,and proposes two different types of pedestrian detection algorithms,and finally merges the detection results of these two algorithms.The final detection result after fusion also successfully solved the problems of inaccurate pedestrian confidence,missed detection and false detection,and has a high application prospect.