Research On Lane Detection Algorithm Based On Infrared Image

Lane detection technology is an important technology in the field of environmental perception of autonomous driving.Lane detection provides an important basis for lane keeping,adaptive cruise and other behaviors of vehicles.Traditional lane detection technology often deals with a single scene and is difficult to adapt to the complex and changeable real driving scene.At present,lane detection technology is mainly based on visible light field.In some complex driving scenes(such as poor lighting conditions at night,low visibility in haze weather,shadow generated by buildings and trees blocking lane lines,other vehicles blocking lane lines,lane line wear,etc.),relevant technologies lack corresponding means.Therefore,based on this background,this dissertation will mainly study lane detection technology used in the infrared field.The main work is as follows:(1)Currently,few lane datasets in the infrared field can be obtained in the open-source platform.The infrared night vision device was installed on the vehicle horizontally,and the vehicle was driven to collect lane images on urban roads and roads around the city.Then,the dataset images were selected,classified and marked.A point-labeling method similar to the visible dataset Tu Simple is adopted in the dataset.(2)An improved ENet multi-task instance segmentation network structure with a recurrent feature shift aggregator is proposed to ensure sufficient acceptance domain for semantic segmentation without sacrificing spatial resolution.The prior shapes of lane lines are used to capture inter-row and inter-column pixel spatial relationships.The addition of instance segmentation branches to the original semantic segmentation ensures the accuracy of subsequent individual classification of lane lines.The experimental validation shows the feasibility of lane line detection within the infrared domain,which shows good detection power in the unfavorable environment of night,shadow and hazy weather,the accuracy and real-time performance of the model are improved than the original model,the model is also able to deal with some bad conditions of vehicle occlusion and lane wear in a targeted manner.(3)An improved Transformer-based lane detection method is proposed to improve the pixel-level segmentation idea of the feature aggregation and multi-task instance segmentation network for the infrared scene when cracks,repair marks,stains and other similar general lane elongated shape structures appear on the road surface,which can produce false lane detection situations,and also to further improve the response speed of the algorithm model by designing a backbone network,in addition The non-local interactions are modeled using Transformer’s self-attentive mechanism to capture the lane slenderness structure and global contextual information,and the lane parameters are directly predicted by designing a suitable feedforward network.Due to the characteristic of lane partial parameter sharing,the obscured or badly worn lanes are able to predict the parameters directly through the surrounding lanes without the interference of road traces.The algorithm improves the detection accuracy compared to the sequential algorithm,and the model response speed reaches 98 FPS.In this study,extensive experiments were conducted on a self-built infrared dataset and also re-run on the standard Tu Simple lane dataset in the visible domain,both of which verified the accuracy and real-time performance of the two lane detection algorithms proposed in this dissertation.