一种基于增强校准策略的自动驾驶多任务分割网络

A Multi-Task Segmentation Network Based on an Enhancement-Calibration Strategy for Autonomous Driving

摘要: 在自动驾驶中，可行驶区域检测与车道线检测是两项语义上紧密相关的关键视觉任务。然而多数方法仍将两项任务独立处理，未能有效利用任务间的关联，此外，大规模网络结构庞大、计算量高，难以满足实时嵌入式车载系统的应用需求。本文提出一种基于增强–校准策略的轻量化多任务语义分割网络GMSANet (Grouped Multi-Scale Attention Network)，用于同时处理可行驶区域检测与车道线检测任务。模型以GSConv-ESP (Grouped Shuffle Convolution-Efficient Spatial Pyramid)结构为编码器基础，通过增强–校准设计，在保持高精度的同时降低模型复杂度。网络引入分组多尺度注意力(Grouped Multi-Scale Attention, GMSA)模块，以分组条带卷积提升对方向性与关键区域的响应；同时提出多尺度动态矩形自校准(Multi-scale Dynamic Rectangular Self-Calibration Module, MD-RCM)模块，通过多尺度感受野实现对目标区域的自适应校准。实验结果表明，在BDD100K数据集上，GMSANet以仅2.9M参数和6.45 G FLOPs的复杂度，实现了92.8%的可行驶区域mIoU、85.1%的车道线准确率与34.0%的车道线IoU，性能优于YOLOP和A-YOLOM等轻量化方法。模型推理速度可达55 FPS，具备良好的实时性与嵌入式部署潜力。

Abstract: Drivable area detection and lane line detection are two related perception tasks in autonomous driving. Many prior works treat them as two separate problems and do not explicitly model their shared features. In many cases, the required computation and model size exceed what typical embedded platforms can support in real time. This paper introduces the Grouped Multi-Scale Attention Network (GMSANet), a lightweight multi-task semantic segmentation network based on an enhancement-calibration strategy, designed for drivable area and lane line detection. The model is built upon the GSConv-ESP (Grouped Shuffle Convolution-Efficient Spatial Pyramid) encoder and uses an enhancement-calibration design, which reduces model complexity while preserving high segmentation accuracy. We integrate a Grouped Multi-Scale Attention (GMSA) module into the network. GMSA applies grouped strip convolutions at multiple scales, which increases the sensitivity to directional features and key regions. Also, we introduce a Multi-scale Dynamic Rectangular Self-Calibration Module (MD-RCM) that calibrates target regions by adjusting receptive fields across multiple scales. Experimental results on the BDD100K dataset show that GMSANet, with only 2.9 M parameters and 6.45 G FLOPs, achieves 92.8% mIoU for drivable area segmentation, 85.1% LaneAccuracy, and 34.0% LaneIoU, outperforming lightweight models such as YOLOP and A-YOLOM. The model further achieves an inference speed of 55 FPS, demonstrating strong real-time capability and suitability for embedded deployment.