Research On Road Detection Under Rainfall Environment

With the development of computer vision,image-based applications have been widely used in daily life and social production.However,current image algorithms are usually based on idealized scenes such as indoor or good weather,without considering the impact of bad weather on imaging.Rainfall is the most common kind of severe weather,especially in outdoor scenes,which has a great impact on the imaging quality.The images taken in rainy days have the problems of rain marks,rain lines and occlusion,which makes it difficult to distinguish the objects in the image.Therefore,in the field of computer vision,it is of great significance to restore the rainfall image with high quality.Rainfall has a great impact on the outdoor work system.Among them,the field of automatic driving with high safety requirements urgently needs to solve the reliability of image algorithm in rainy scenes.The influence of rainfall on vehicle image acquisition may lead to errors in vehicle decision system and safety accidents.In the field of automatic driving,road detection is the most basic task.The road image under rainfall is blocked by rain lines and it is difficult to accurately get the passable area.Therefore,in order to better improve the reliability of outdoor visual tasks in rainfall environment,especially the basic tasks such as road detection in the field of automatic driving,this paper carries out modeling and representation according to the principle of rainfall imaging,proposes an image rain removal recovery algorithm with high reliability and real-time,and applies it to the actual road detection task to improve the safety of automatic driving.Based on the above background,this paper carries out the following research on image rain removal and road detection in rainy scenes:(1)The imaging characteristics of raindrops are studied,and an appropriate model is used to describe the imaging process of raindrops.MPN(Multi-scale Progressive Restoration Network)is proposed for rain remove and recovery.The non-local information is sensed by dilated convolution and multi-scale fusion mechanism,and the mapping process from rain images to rain removal images is transformed into a learning process of rain map residuals.The recurrent unit is introduced to connect different stages,and the rain removal process in the previous stage is used to guide the rain removal recovery in the subsequent stage.The experiment proves that the proposed method has good de-rain effect and image recovery effect.(2)Based on MPN rain removal model,improvement strategies were proposed to obtain better rain removal effect,and GAN model was used to remove rain streaks to obtain better recovery quality.In order to make the network pay more attention to the distribution area of rain streaks and the occluded background,a parallel attention module and a global attention fusion mechanism are added to the generator.At the same time,in order to improve the discriminator ability,the discriminator structure guided by attention prior was designed to improve the rain removal ability of the generator synchronously.And then,in order to better distinguish the difference between background high-frequency details and rain streaks,Laplace edge perception is introduced to improve the restoration effect of background details and make the generated image background more clear.The experimental results show that the improved MPN-ACG method is superior,and the rain removal performance is obviously improved.(3)The purpose of the rain removal task is to facilitate the development of subsequent image tasks.Therefore,we first explored the performance of mainstream road detection algorithms in rainless and rainy scenes,and embed the proposed improved rain removal algorithm MPN-ACG into the road detection task to achieve the whole process from rain removal preprocessing to road detection,boundary fitting,and then vehicle coordinate conversion.