Research On Pedestrian Detection Technology Based On Deep Learning

With the development of modern science and artificial intelligence,pedestrian detection technology is favored by many researchers in the field of computer vision.Pedestrian detection technology is the basis of gait analysis,pedestrian tracking,human action recognition and other research.It is widely used in monitoring security,vehicle-assisted driving,virtual reality and intelligent robots.In recent years,deep learning has made breakthroughs in all aspects,including more accurate and rapid target detection,and greatly improved detection performance in pedestrian detection.Compared with traditional pedestrian detection methods,deep learning more robust.But there is still a problem of low detection accuracy in occlusion,multi-scale and other environments.This paper mainly builds a more accurate pedestrian detection model based on deep learning.First,the paper constructs a pedestrian detection model based on a faster regional convolutional neural network.Considering that there are many low-resolution samples in the still image during the shooting process,the paper uses the histogram equalization algorithm to preprocess and enhance the image effect.Since Faster R-CNN is for multi-target detection,the paper aims to extract the proposals by modifying the RPN anchor box for the goal of pedestrians.Aiming at the problem of occlusion,the pedestrian detection model of Faster R-CNN is assisted by fusing convolution multi-layer features,setting repulsion loss function and improving non-maximum suppression algorithm.The effects of various factors on the detection performance of faster regional convolutional neural networks are studied by contrast experiments.The results achieve an ideal detection effect on the INRIA dataset and some Caltech datasets.Then,in order to solve the problem of pedestrian multi-scale in complex background environment,the paper gives the context refinement algorithm and combines with RPN to make RPN generate more accurate pedestrian regions.In addition,based on the principle of feature pyramid network,the high-level information is transmitted to the low-level through the way of up-sampling,which is fused with the low-level features and predicted independently by using different feature layers.It is integrated into RPN and Fast R-CNN networks to effectively solve pedestrian multi-scale problems.Experiments in INRIA,Caltech and some supplementary datasets show that the proposed algorithm achieves high accuracy and proves the effectiveness of the proposed method.