Research On Lane Detection Algorithm Based On Deep Learning

In recent years,China’s road transportation system has been continuously improved,and the intelligent driving system has become increasingly mature.At the same time,lane line detection,as an important part of intelligent driving technology,provides basic environmental information for subsequent decision-making and response of vehicles.However,the current lane line detection method is only suitable for simple scenarios,and it is difficult to balance the accuracy and speed of detection in the face of complex road traffic environment.Therefore,in order to further improve the safety of intelligent driving,it is necessary to conduct more indepth research on lane line detection.In this paper,the two existing models are improved and integrated,a strengthened spatial information association module is designed,and a lane line detection model is constructed based on this module,and finally the effect of the module and the overall model is verified by experiments.In this paper,the attention mechanism model is first improved,so that it can give weights to the spatial domain and channel domain of the feature map at the same time,and expand the receptive field of the model.The slice convolution model is improved,and the original sequential convolution structure is changed into parallel convolution,which greatly reduces the running time of the model while ensuring that the information of the feature map can be transmitted row by row and column.The two improved models are fused to design and strengthen the spatial information association module,which can effectively improve the model’s inference ability for invisible lane lines and run fast.Secondly,based on the enhanced spatial information association module designed in this paper,a complete lane line detection network model is constructed.The model adopts the series structure of encoder-enhanced spatial information association module-decoder,in which the encoder adopts the main structure of ResNet to remove its full connection layer,which plays the role of downsampling the input image to form a feature map.The encoder uses a combination of linear interpolation and transpose convolution for upsampling,restores the feature map to the input image,and adds a branch for judging the existence of lane lines to jointly complete the task of detecting lane lines.Finally,in order to verify the effect of the lane line detection model and the enhanced spatial information association module proposed in this paper,a horizontal comparison experiment and an ablation experiment are designed.Two public datasets,CULane and TuSimple,were used to train and test the model.Cross-sectional comparison experiments show that the lane line detection model in this paper does not improve significantly on the simple scene(TuSimple),but the accuracy(F1)in the complex scene(CULane)is 72.3,which is 15.2%higher than the DenseCRF model and only 1.8%lower than the SCNN model,but the running time is only 56.2%of that of SCNN.The ablation experiment shows that in complex scenarios,the detection accuracy(F1)of the model with the enhanced spatial information association module is improved by 17.8%compared with the model without the added model,and the operation time is only increased by 1%～2%.In simple scenarios,the improvement effect is limited,5.8%.In conclusion,the effectiveness of the proposed module is verified,and the lane line detection model designed based on this module has stronger robustness and faster detection speed.