Semantic Segmentation Of Road Marking Based On Deep Convolutional Networks

Autonomous driving technology is considered to be a disruptive technology with huge economic and social value.Research on autonomous driving technology focuses on three aspects: perception,decision-making,and control.Among them,the task of semantic segmentation of road marking is the technical bottleneck of the current perception module,and has become a hot research direction.Since the traditional semantic segmentation technology uses artificially designed features,it is more suitable for applications with simple scenes.However,the changes of road scenes in real environments are very complicated,and they often face problems such as light changes,road marking wear and occlusion,which leads to a significant reduction in the robustness of related algorithms.In recent years,convolutional neural networks have been introduced into the field of image segmentation as a new method,and have achieved a series of excellent performances.This paper applied convolutional neural network to road marking semantic segmentation tasks to obtain accurate and real-time semantic segmentation effects in complex and changing road scenes.The specific research contents are as follows:①Based on the encoder-decoder architecture of the convolutional neural network,introduces the porous spatial pyramid pooling structure and proposes the MarkNet network.Focus on analysis and research on the impact of category imbalance problem and loss function selection problem on model performance,and on this basis,comprehensively apply category rebalance weight and joint loss function to improve MarkNet’s segmentation performance.Finally,experiments prove that MarkNet can realize real-time and high-precision road marking semantic segmentation tasks in complex road scenes.②Draws on the idea of LaneNet’s branch network,introduces a binary branch network into the MarkNet network,and proposes a MarkNet + network.In order to further improve the segmentation performance of the MarkNet network and reduce the problem of blurring or sticking of object edges in the segmentation results,a binary branch network is introduced to learn the location segmentation of the target.Experiments prove that MarkNet + has improved the segmentation performance of different types of road markings compared to MarkNet.Its segmentation performance can meet the accuracy requirements in complex environments,and its speed basically meets the real-time requirements.The experimental results on the ApolloScape driveless dataset show that in the face of complex and changing road scenarios,the MarkNet + network proposed in this paper can achieve accurate and real-time semantic segmentation of road markings.