Research On Lane Line Detection In Complex Environment Based On Deep Learning

Lane line detection is the key technology to realize driverless and intelligent vehicle aided driving,which can provide accurate road information for vehicles.However,the environment is more complicated in the actual road scene of the vehicle.The lane line is disturbed by changes in light,shadows,occlusion,and road surface damage,which makes it more difficult to detect.To this end,this paper takes lane line detection in complex environments as the goal,studies the lane line feature point detection method under interference conditions,the distance measurement method at the lane line dashed-solid line transition point,and the stable continuous tracking of lane line.The specific content of the paper is as follows:(1)Aiming at the problem that lane line feature points cannot be accurately detected due to blurring,shadow occlusion or breakage,etc.,a lane line feature point detection method in a complex environment based on the fusion of a non-local neural network module and an encoder-decoder structure is proposed.Improve the feature detection method by combining the non-local deep neural network module with the encoder-decoder module,recover the image size by upsampling and deconvolution,and finally use the Softmax layer to segment the lane image.Experiments show that the improved lane segment recognition algorithm based on the improved semantic segmentation network can achieve 90% accuracy when the lane line feature points are damaged,blurred and occluded by shadow(2)In order to solve the problem of ranging between the dotted line and the solid line and the current vehicle in the complicated environment,a multi-task neural network based method is proposed.Based on the research of lane line feature point detection in a complex environment,a multi-task neural network structure is constructed.After the original coding structure,a branch structure is added to extract different types of lane line types in different regions of interest.Location characteristics.Add radial distortion correction to camera calibration.The experimental results show that the ranging error at the type conversion point can be controlled within 5%.(3)Aiming at the problem of low accuracy of lane line fitting due to vehicle occlusion during lane line tracking,a lane line tracking model that incorporates three-dimensional position information of vehicles is proposed.Based on the traditional Kalman filter-based lane line tracking state matrix,the vehicle's three-dimensional position information is added,and the lane line tracking state matrix is extended to improve the effect of lane line tracking detection.The experimental results show that the average absolute value error of the lane line tracking model fused with the three-dimensional position of the vehicle is reduced by 7%,which has a good effect.In order to verify the effectiveness of the proposed method,a multi-tasking lane line detection experimental platform based on the Tensorflow framework is designed and constructed.The vehicle driving image data set collected in actual scenes and Tucson data set are used to perform lane line feature points.Detection and classification experiments.Experiments prove that the method proposed in this paper has higher accuracy for lane line feature point location and type recognition,and has higher practical value.