Lane Detection In Nighttime Based On Deep Learning

As one of the basic task of automatic driving,lane detection has a long history and a wide range of applications.Nowadays,existing deep learning-based lane detection methods rely on the powerful feature learning capabilities of convolutional neural networks(CNNs),which can extract better discriminating features between lanes in daytime scenes.But due to the low contrast of night images,it is difficult for CNNs to identify lane.Therefore,how to improve the lane feature extraction structure and increase the contrast of night images are two methods to help CNNs extract discriminative features,thereby improving the performance of night lane detection.In addition,existing publicly available datasets of lane detection have a variety of scenarios,but their distribution is uneven.For example,the number of nighttime scenes is less than daytime scenes,which motivates us to utilize daytime scenes to synthesize nighttime scenes,thereby increasing the nighttime scenes during training.This article explores lane detection in the night environment based on deep learning,and the main work is as follows:(1)Design a parallel multi-scale convolution module,which can be applied to the existing backbone network to enhance the learning of lane features in the night environment.Specifically,we adopt a multi-scale convolution on the deep feature maps to obtain rich receptive fields,so as to learn robust features.And the inference efficiency of the module is ensured by the parallel mode of loop multi-step length,and the information transmission in a single direction is completed through faster iterations.In addition,the overall network architecture uses feature fusion to compensate for the local information of deep features and further optimizes the model’s ability to extract lane features.Compared with other methods,our method obtains better performance on the major datasets.(2)Design a lane detection method based on unsupervised low-light image enhancement.Specifically,we use an unsupervised curve estimation method to enhance the low-light image in the night environment,thereby improving the contrast of the night image.Therefore,the lane in the low-light image will be enhanced,which makes it easier for CNNs to extract discriminative features for effective lane detection.Moreover,we also design a night lane attention mechanism to further enhance the network’s ability to learn the discriminative features of lanes,thereby improving the performance of lane detection.We conduct the experiments on the CULane and BDD100 k datasets.Experimental results further prove that our method can achieve the state-of-the-art on these two datasets.(3)Propose a dual-stream network,which consists of a daytime branch,a nighttime branch,and a lane prediction module.The dual-stream network takes a pair of daytime scene and synthesized nighttime scene as input,and predicts the lanes of the daytime and nighttime scenes as output.Inspired by image translation,we utilize an image translation network to translate daytime scenes into nighttime scenes.In this way,we can obtain the image pairs of daytime scene and nighttime scene.In order to improve the performance of lane detection,we design an embedded loss to help the night branch learn discriminative features from the day branch.In addition,in order to strength the feature extraction of lane structure,we also propose a streakaware pooling module to aggregate local contextual information and further improve the performance of lane detection in the night environment.Extensive experiments on the nighttime scene of the CULane dataset demonstrates the effectiveness of our method.