L2-LiteSeg: A Real-Time Semantic Segmentation Method for End-to-End Autonomous Driving

The vehicle-borne optical imaging sensor system cannot accurately and quickly receive information about the driving environment, which may lead to the Autonomous Driving System (ADS) being unable to make accurate and rapid decisions, thus bringing potential safety hazards to autonomous driving. In order to solve the above problems, we propose a real-time semantic segmentation method for end-to-end autonomous driving. This method improves the practicability and reliability of autonomous driving technology by improving the real-time and accuracy of the semantic segmentation method. First, in order to solve the problem that existing semantic segmentation methods cannot meet the requirements of vehicle-borne embedded terminals for real-time and segmentation accuracy of semantic segmentation methods due to complex network structure, difficulties in extracting useful features in complex autonomous driving environments, high training hardware configuration, we propose a real-time semantic segmentation network model L2-LiteSeg, it solves the problems of the low extraction efficiency of feature and parameter redundancy by optimizing the configuration of Unified Attention Fusion Module (UAFM) modules and the network structure of the PP-LiteSeg model. On semantic segmentation datasets for autonomous driving and CamVid, L2-LiteSeg only uses 0.82M parameters, the semantic segmentation precision reaches 73.1%, 67.7% mIoU, and 67.7% mIoU respectively, and the reasoning speed reaches 73.68 Frames Per Second (FPS), 85.16 FPS, and 73.68 FPS respectively. This shows that the L2-LiteSeg method has achieved a good balance in the three indicators of parameter, accuracy, and reasoning speed and can meet the real-time and accuracy requirements of semantic segmentation by the vehicle-borne embedded terminals.