Multi Scene Lane Detection Based On Image Processing

Along with the advent of artificial intelligence and the presence of Internet of Vehicles and 5G networks,intelligent driving vehicles have sprung up.Safety issues of intelligent driving are increasing stand out since complex environment of roads and impaired traffic signs.Information perception of driving environment is extremely crucial for safety driving of intelligent vehicles.Advanced driver assistance systems(ADAS)have remarkable advantages in accessing information inside and outside of the vehicle to keep driving safety.As an essential part of ADAS,lane detection plays an important role in departure warning,lane keeping and path planning and decision-making.The diurnal variation of light and damaged roads in complex traffic scenarios usually act as serious interfering factors for lane detection.Meanwhile,ADAS has high requirements for real-time lane line detection,which makes improving the speed of detection an urgent problem.Thus,it has become a challenging and application-worthy topic to design an accurate real-time lane line detection model.In response to the above problems,an image processing-based lane line detection model for multiple scenes is proposed,which works as follows:(1)We put forward a new loss function,auxiliary loss,for the gradient disappearance problem that appears in back propagation,which adds an auxiliary training branch after the semantic segmentation encoder,then modify the calculation of cross-entropy loss by back-propagating the updated parameters.At the same time,the lane predictor is added in parallel with the decoder to compensate for the poor pixel-level semantic segmentation accuracy.The semantic segmentation loss,auxiliary loss,and lane prediction loss are weighted and summed to work in concert for back propagation.(2)For the problems of poor real-time and accuracy of lane line detection,we propose an lane line detection model with improved attention pyramid and focus loss.Our model exploits the advantage of weak bottleneck residual module in the encoder in reducing the number of model parameters.We improve the weak bottleneck residual module using jump coupling of residuals to solve the problem of information loss between adjacent asymmetric convolution layers in the residual block.The attention pyramid module in the decoder is improved to form two small pyramids,and the feature pyramid attention is added on the high-level output to extract rich contextual features.Meanwhile,to solve the problem of unbalanced distribution of lane line pixels and background pixels caused by sparse lane line information,an improved focus loss sample equalizer is proposed to accelerate the convergence of the loss by adjusting the settings of hyper parameters.As a proof of the working of our model,we validate it on the lane line dataset CULane in terms of comparing several metrics such as F1,FP,run time,and number of parameters.