Structure-aware Lane Marking Detection Methods

Lane marking detection is a fundamental task,which serves as an important prerequisite for automatic driving or driver-assistance systems.Early works of lane marking detection mainly used hand-craft features,but they are limited to daytime highway scenario.Later,with the rapid development of deep learning,Deep Convolutional Neural Network(DCNN)was applied to lane marking detection task.And deeplearning-based lane marking detection methods greatly improved on both accuracy and robust.However,under complex and uncontrollable driving road environment,such as crowd,poor illumination,etc.,lane marking detection task is still challenging.To solve this problem,existing lane marking detection methods used vanishing point as guidance or enlarged the receptive field,but still failed to aware the long discontinuous appearances of lane markings.In this paper,two different methods were presented to detect lane markings in a complex environment by analyzing their structure information.One of them is a structure-aware lane marking detection network(SALMNet),which contains two presented modules.Firstly,a semantic-guided channel attention(SGCA)module is designed,which selects the low-level features of a deep convolutional neural network by taking the high-level semantic features as guidance.Secondly,a pyramid deformable convolution(PDC)module is presented,which enlarges the receptive fields and captures the complex structures of lane markings by applying deformable convolutions on multiple feature maps with different scales.Hence,the SALMNet can better reduce false detection and enhance lane marking structures simultaneously.The experimental results on three benchmark datasets for lane marking detection show that the SALMNet outperforms other methods on all the benchmark datasets.A bidirectional structure fusion lane detection network is presented to aggregate multilevel features by properly designed hierarchical aggregated deformable convolution(HADC)module.HADC module learns lane marking structures by deformable convolution,and feeds the features to lower stages,and then passes low level details upward.Multiple HADC are stacked to fully fuse features from different stages of DCNN.HADC module is verified on three different public datasets and the experimental results indicate that HADC module promotes lightweight DCNN models greatly.In summary,our major contributions of this work are as follows:1)A structure-aware lane marking detection network(SALMNet)is developed to aware lane markings structures,which contains a semantic-guided channel attention module to enhance the features on the lane marking regions and suppress noise from the background by selecting feature maps for lane making detection and a pyramid deformable convolution module to enlarge the receptive fields for DCNN and to obtain more structure information of lane markings.The SALMNet is evaluated on three public datasets and further compared with stateof-the-art methods.Experimental results show that the SALMNet performable favorably against the state-of-the-art methods on all benchmark datasets.2)A bidirectional structure fusion lane marking detection network is presented to aggregate the structure information of lane markings by stacking multiple hierarchical aggregated deformable convolution module which propagates structure information captured by deformable convolution downwards and deliveries low-level detail information of a DCNN upwards.The model is compared with the SALMNet as well as other state-of-the-art methods on three benchmark datasets,and experiment results show that under lightweight feature extraction networks the model performs well.