A Lane Detection Algorithm Based On Deep Learning And Research On Lane Departure Warning

For intelligent transportation,autonomous driving and advanced driver assistance systems can reduce traffic accident rates and improve travel comfort.Lane detection is an important part of system perception,and lane departure warning is a key technology to ensure driving safety.The lane detection algorithm based on traditional vision cannot deal well with the lane detection problem in complex scenes such as crowded,night,and absence.The lane detection algorithm based on deep learning has strong adaptability and high detection accuracy.In recent years,it has received more and more research and attention.The lane detection algorithm based on deep learning and lane departure warning system are studied in this thesis.The speed and accuracy of lane detection are improved.The robustness and real-time performance of the lane departure warning system are enhanced.The specific work is as follows:（1）A multi-branch lane detection algorithm based on the RepVGG network is proposed.On one hand,the RepVGG backbone network is improved.The attention mechanism SENet module is introduced in the backbone network to enhance the feature extraction of important information of the lane lines.The training state model is decoupled with the structural re-parameterization method to obtain faster inference state model having the same performance.On the other,a multi-branch lane detection method based on the classification in the row direction is proposed.A row-by-row detection branch is added behind the backbone network for reducing the amount of calculation and realizing the detection of occluded or defective lane lines.An offset compensation branch is designed,which refines the predicted location coordinates in the local range in the horizontal direction and recovers lane details.Parallel separable auxiliary segmentation branch is added to model local features for improving detection accuracy.The algorithm is tested on the public lane detection data set CULane,and compared with the UFAST18 algorithm,which is the fastest lane detection algorithm at present.The result shows that inference speed is increased by 19%,and model size is reduced by 12%,and the evaluation index F₁-measure increases from 68.4 to 70.2.And the algorithm inference speed is 4 times that of the SAD algorithm and 40 times that of the SCNN algorithm.（2）Research on lane departure warning model.The calibration experiment of the vehicle camera is carried out,and the internal and external parameters of the camera are obtained.Based on the results of the lane detection,the lane lines are fitted by the least squares algorithm.A lane tracking module based on Kalman filter is introduced to correct the lane lines with poor detection effect.The TLC（Time to Lane Crossing）departure warning algorithm based on the time dimension and the CCP（Car’s Current Position）departure warning algorithm based on the space dimension are studied,the latest warning boundary is introduced,and a lane departure joint warning model based on the latest warning boundary is constructed.（3）Experimental test and analysis.The software test platform is developed through GUI programming,and real vehicle data acquisition and experimental test verification are carried out under urban road conditions.In most scenarios,such as normal,crowded,and tunnels,the number and shape of lane detected are consistent with the actual situation.In common scenarios,the lane missed detection rate is between 0 and 10%,which indicates that the lane detection algorithm based on the RepVGG network has strong generalization ability.The average hit rate of the lane departure warning test is 91.54%,the average false alarm rate is 0.90%,and the average test time per frame is 33～35 milliseconds.By the software platform test,the lane lines predicted are clear,the warning result of the warning interface is correct.