Research On Road Condition Recognition Technology Based On Near Infrared Imaging

Road condition recognition is a key link in the growing self-driving technology.Its purpose is to identify and locate obstacles from complex driving environment,and provide preliminary information for subsequent analysis and decision-making.In road condition recognition,the detection of pedestrians and vehicles has received extensive attention.In this thesis,the self-driving vehicles is taken as the research object,we mainly focus on the pedestrians and vehicles detection in the driving environment.In current road condition recognition technology,the acquisition of driving environment information is mainly based on visible light imaging system.However,this system is greatly affected by the weather and light.It is difficult to get clear driving image in bad weather and poor light environment,which brings serious interference to the pedestrians and vehicles detection.To settle the aforementioned problem,the pedestrian and vehicle detection under poor lighting conditions is taken as the research direction.The main contributions of this thesis are as follows:1)A relatively complete driving environment database is constructed with the near infrared imaging system.In this thesis,the active near-infrared imaging technology is used to effectively solve the problem of clear imaging under poor illumination conditions.In order to train and test the algorithm model,the driving environment information in many different scenarios is collected,and a complete pedestrian-vehicle data set is constructed by labeling this information.2)An object candidate region extraction algorithm based on image segmentation is proposed.In near-infrared images,objects reflecting near-infrared rays are obviously brighter than the surrounding background.In view of this characteristic of near infrared image,a algorithm for culling the background region in the image by a threshold segmentation algorithm and reducing the range of images that need to be traversed to obtain the possible area of the target is proposed.In the threshold segmentation algorithm,by using the proposed decomposition calculation,the multi-dimensional gray information of the image is utilized without making the algorithm time complexity too high.In this thesis,the selective search algorithm is used to obtain the possible locations of pedestrians and vehicles in the image,and its calculation function is adjusted to make it suitable for near infrared images.3)An adaptive scale-aware target detection algorithm is proposed.The distancebetween the target and the imaging system results in a different proportion of the target size on the image.Targets at different spatial scales may exhibit distinctly different characteristics,posing challenges for target detection.In response to this problem,a divide-and-conquer strategy is proposed,in which multiple sub-networks are built in the detection model to detect targets of different scales.By defining a gate function for the target height,the target is sent adaptively to the corresponding subnet for detection.In order to improve the detection speed,a convolution neural network based on full image is constructed,and the detection sub-networks share the convolution feature of the entire image.