Study On Object Detection Of Foggy Road Based On Convolutional Neural Network

With the development of deep learning and driverless technology,the application of road object detection is becoming more and more widespread.Road object detection has high requirements for real-time performance and accuracy.In addition,severe weather conditions such as foggy weather often occur in outdoor environments,which greatly affects the effectiveness of object detection algorithms.Studying the problem of road object detection in foggy environment and improving the accuracy of existing object detection algorithms are of great significance to the development of related industries.Based on convolutional neural network,this thesis studies the object detection algorithm of foggy roads.Based on the YOLOv3 network,this thesis proposes an improved object detection network that combines deep separable convolution and dense residual networks.The network improves the feature extraction module of the YOLOv3 network,and increases the number of convolution modules and the YOLO layer.Compared with the YOLOv3 network,the object detection network has a greater improvement in the m AP@50 evaluation metric.Due to the reduction in the amount of parameters and calculations,the detection speed has also increased.Based on the above work,this thesis also proposes a pre-selection box clustering algorithm based on K-means ++ algorithm and GIOU,which improves the selection of pre-selection boxes and further improves the accuracy of detection.In the experimental stage,this thesis extracted 20,000 pictures from the BDD100 K data set as the data set.The improved network proposed in this thesis has a m AP@50 evaluation value of 56% on this dataset,which is better than the original YOLOv3 network.In order to solve the problem of low recognition rate in foggy object detection,this thesis proposes an embeddable defogging neural network DefogNet based on the atmospheric scattering model.DefogNet uses the techniques of lightweight models such as inverted convolution and depth separable convolution.In addition,DefogNet references the method of AOD-Net fusion atmospheric scattering model,and regards the atmospheric light and transmission map as a variable K for prediction.In the experimental stage,this thesis firstly processes the Cityscapes dataset with fogging algorithm to obtain the corresponding foggy dataset Cityscapes-fog.The dataset is used to train and test DefogNet and other defogging models.Finally this thesis verifies the superiority of DefogNet.In the method of model fusion,this thesis adopts the direct-tandem method.Fog images are first dehazed by DefogNet.The fog-free images are further detected by D2-YOLOv3 algorithm.In the experimental stage,Cityscapes were used to train the D2-YOLOv3 model,and the fog validation set was used to verify the model’s effect.The detection effect of the D2-YOLOv3 network combined with the defogging model on foggy images is 6% higher than that of the D2-YOLOv3 single model on the m AP@50 evaluation metric.