Research On Traffic Object Detection Method In Foggy Based On YOLOX

With the rapid development of economy and the continuous expansion of cities,people’s demand for living standards and travel is also increasing.With the number of cars increasing year by year,traffic safety accidents are becoming more frequent.In adverse weather(fog,haze,etc.),the driver’s sight is disturbed,and it is difficult to fully grasp the surrounding traffic and road conditions,resulting in driving judgment errors.In particular,the number of traffic accidents under the increasingly severe fog-haze weather is significantly higher than that under the sunny weather.Therefore,adverse weather has become an important factor affecting the safe driving of traffic roads.How to effectively prevent and reduce the occurrence of such problems is of great significance.This dissertation will introduce and study the traffic target detection technology in foggy.The main work and research innovation are as follows:(1)In this dissertation,the traffic target detection and recognition under the foggy weather conditions and related theoretical knowledge are analyzed and discussed in detail,and the methods of how to effectively improve the accuracy of the object detection and recognition process under the fog weather conditions are deeply studied.At the same time,the correlation method is tested,the performance of the correlation method is analyzed,and the key problems needing attention in the process of object detection under fog-haze weather conditions are summarized.(2)Aiming at the problem of too shallow network of AOD-Net dehazing algorithm and the problem that the overall color,brightness and distortion of the image will be affected when processing foggy images,an improved AOD-Net dehazing enhancement method is proposed.In this method,four max-pooling layers are respectively added at the end of the backbone of the dehazing network AOD-Net to extract and amplify the detailed features of the image,and then the features after four bilinear up-sampling are respectively integrated.At the same time,a loss function MS＿SSIM + L1 that can preserve the contrast of the high-frequency region of the original image and the color and brightness of the original image is introduced to replace the original Mean Square Error loss function to obtain the final dehazing image.By comparing the subjective and objective evaluation of the dehazing image,the method improves the dehazing effect of the image.(3)Aiming at the problems of poor visual effect,high noise and low recognizability of the images collected by monitoring in fog weather,a YOLOX dual backbone object detection method based on transfer learning is proposed.The dual backbone feature extraction network structure forms a more powerful backbone feature extraction network by composite connection between adjacent backbone networks.The structure takes the output of each feature level as a part of the input,flows to the parallel stage of the subsequent backbone through composite connection,and merges multiple high-level and low-level features to generate a richer feature representation.At the same time,the output features of the backbone are fed back to the subsequent backbone iteratively as part of the input features in a stage by stage manner,and finally the features of the last backbone,that is,the leading backbone for object detection.In this dissertation,the model parameters pre-trained on the large source domain dataset MSCOCO are migrated to the small target domain dataset RTTS by using the transfer learning technology,and the network parameters are fine-tuned to adapt to the new dataset.The experimental results show that it can significantly improve the final object detection and recognition.