Research On Nighttime Lane Detection Methods

Lane detection technology is a crucial aspect of the intelligent transportation and autonomous driving industries.It aids autonomous vehicles in performing lane-changing maneuvers by accurately detecting the real-time position of lane lines,resulting in enhanced road traffic efficiency and safety.However,low-light environments,such as nighttime scenes,pose significant challenges for detecting lane lines,including insufficient lighting,light source interference,and varying weather conditions.Therefore,this article’s primary focus is to examine lane detection methods specifically designed for nighttime scenarios.The main areas of focus include:Ⅰ.The backbone network is ERFNet,a lightweight semantic segmentation network,which is used due to the computational constraints of onboard systems.In order to address the problem of inadequate prediction capability of the backbone network,a lane prediction branch is added after the encoder to assist in determining the presence of lane lines.To enhance the network’s ability to extract global features and improve the accuracy of lane detection in scenes with occlusion,shadows,and other challenging scenarios,a self-attention mechanism is being introduced,which is inspired by the slender spatial structure of lane lines.Experimental results have shown that SA-ERFNet achieves improved accuracy in lane detection,particularly in challenging scenarios involving occlusions and shadows,while maintaining real-time performance.Ⅱ.In order to tackle the problem of diminished accuracy in lane detection for SA-ERFNet when operating in nighttime environments,a solution is presented wherein an image enhancement technique known as Zero＿LE is integrated into the front-end of the lane detection process.It learns high-order enhancement curve parameters to enhance the quality of nighttime driving images.RGB attention map is added to the network input to enhance the brightness in dark areas,DO-Conv is used to replace standard convolution to reduce model computation,and a Channel-Spatial Dual Attention Module(CSDAM)is designed to suppress noise and reduce color distortion by focusing on the noise positions in the image and learning different channel color features.Experimental results show that Zero＿LE not only produces well-exposed and naturally colored enhanced images,but also improves the accuracy of lane detection in nighttime scenes.Ⅲ.By drawing on the multi-exposure fusion low-light image enhancement method,further improvement is made to the Zero＿LE method to achieve a more natural visual effect for nighttime driving images and enhance the accuracy of nighttime lane detection.For a single low-light lane image,an adaptive gamma correction is used to generate a clearer and more natural weakly exposed image,while the Zero＿LE method is used to generate a well-exposed enhanced image.An exposure interpolation method is then used to generate a virtual image with moderate exposure.Finally,the fusion weights are characterized using three measurement factors: contrast,saturation,and good exposure,for image fusion.Experimental results demonstrate that SMF can effectively improve brightness and contrast while avoiding color distortion,leading to significant improvements in lane detection accuracy.