Research On Foggy Vehicle Detection Algorithm Based On Deep Learning

In recent years,the widespread use of smart driving technology has provided great convenience for people’s daily travel.However,with the increasing emphasis on traffic safety,vehicle detection technology in bad weather has become more important.Vehicle vision sensors capture blurred images and indistinguishable targets in haze scenes,a situation that causes degradation in the performance of deep learning algorithms to obtain desirable predictions.At present,existing lightweight defogging algorithms are dedicated to improving the model inference speed while it is difficult to take into account the actual defogging requirements,mainly in the shallow network structure and insufficient feature information extraction of lightweight defogging algorithms,which leads to poor defogging effect of images and makes it difficult to provide clear and fog-free images for subsequent target detection tasks;secondly,existing detection algorithms often use a tandem feature extraction network,which is affected by foggy days,and the extracted feature information has more noise interference,resulting in lower detection accuracy of the model.In order to solve the above two problems,this paper investigates a lightweight fog removal algorithm based on deep learning and a two-stage target detection model with the goal of improving the vehicle detection accuracy in foggy scenes,and the main research is as follows:An improved lightweight image defogging algorithm is proposed for the problem of incomplete defogging and chromatic aberration of AOD-Net defogging algorithm.Firstly,the first two convolutional layers of AOD-Net are position normalized,and the extracted moment information is input into the subsequent network layers for affine transformation to improve the data distribution in the original network and optimize the convergence capability of the network.Second,an attention module is introduced to extract feature information using a multi-scale convolutional kernel,weight fusion features and adjust the weights of network channels to suppress redundant information and improve the defogging quality of the model.Comparing and analyzing with existing lightweight defogging algorithms on the publicly available dataset RESIDE,the experimental results show that the defogging quality of the improved AOD-Net algorithm is better than that of defogging algorithms such as DCP,CAP and Dehaze-Net.Compared with the original AOD-Net,the peak signal-to-noise ratio PSNR of the optimized algorithm is improved by 2.71 dB,and the structural similarity SSIM reaches 0.94,which effectively improves the image defogging ability of the network and can provide clear defogged images for the subsequent detection tasks,further improving the detection accuracy of the model.A target detection algorithm(High Resolution Cascade R-CNN,HR-Cascade R-CNN)incorporating high resolution network is proposed to address the problem of foggy scenes where targets are difficult to distinguish and other problems that lead to wrong and missed detections.The high-resolution network is used as the feature extraction network of Cascade R-CNN,and the sub-networks with different resolutions are connected in parallel to extract multi-scale feature information,reduce the information loss in the downsampling process,and enhance the semantic information representation of the target;the original Smooth L1 loss function is replaced by the CIOU loss function,and the penalty term is introduced to measure the width-to-height ratio between the real frame and the detection frame.Finally,SoftNMS is used to improve the candidate frame selection mechanism,reduce the leakage rate in the case of vehicle occlusion,and improve the detection capability of the network.The experimental results on real foggy weather dataset show that HR-Cascade R-CNN can effectively detect the target information of vehicles and pedestrians in foggy weather scenes.Compared with the original Cascade R-CNN,the average detection accuracy of HR-Cascade R-CNN combined with the improved AOD-Net algorithm before and after defogging is improved by 5.9%and 6.9%,respectively,with good detection performance.