| 摘 要: 车道线检测作为自动驾驶领域的一项关键技术，其精度及计算复杂度对车辆的定位和行驶安全具有决定性影响。然而目前检测精度较高的CLRNet模型存在计算复杂度高、难以部署于轻量级终端设备等不足。基于此，提出一种轻量化CLRNet的车道线检测算法：用MobileNetV3-Large网络替代原模型中的ResNet18特征提取网络，降低计算复杂度；并在特征提取网络中引入坐标注意力机制，增强空间特征表征能力。实验结果表明，改进后的模型在Tusimple公开数据集上达到96.57%的检测精度，较原模型仅降低0.27个百分点，模型参数量2.41M，较原模型的11.77M减少79.52%，计算复杂度降至2.80 GFLOPs，并实现210 帧/秒的实时推理性能。在基本维持高精度的前提下，该模型显著降低了模型参数量和计算复杂度，适用于移动终端设备等环境下的车道线检测任务。 |
| 关键词: 车道线检测 轻量化CLRNet MobileNetV3-Large 坐标注意力 |
|
中图分类号: TP391.41
文献标识码:
|
| 基金项目: 南京邮电大学横向科研项目（2024外403） |
|
| A lightweight lane detection method based on CLRNet algorithm |
|
CHEN Shuhui1, ZHOU Zhengkang2, TANG Jiashan1
|
1.College of Science, Nanjing University of Posts and Telecommunications;2.Nanjing Urban Construction Tunnel and Bridge Intelligent Management Company Limited
|
| Abstract: Lane detection, as a pivotal technology in autonomous driving, critically influences vehicle positioning and driving safety. However, the current CLRNet model, despite its high detection accuracy, suffers from high computational complexity and poses challenges for deployment on resource-constrained devices. Based on this analysis, a lightweight algorithm for CLRNet lane detection algorithm was proposed. Firstly, replacing the original ResNet18 backbone with MobileNetV3-Large to substantially reduce computational complexity. Secondly, integrating a coordinate attention mechanism into the feature extraction network to strengthen spatial feature representation. Experimental results show that the improved model achieves 96.57% detection accuracy on the Tusimple public dataset, representing only a marginal degradation of 0.27 percentage points from the baseline, while achieving a 79.52% reduction in parameter count from 11.77M to 2.41M. Furthermore, the computational complexity is reduced to 2.80 GFLOPs while achieving an inference speed of 210 frames per second. This design achieves an optimal balance between precision and efficiency, significantly alleviating computational and storage requirements while maintaining competitive accuracy, making it suitable for lane detection task in resource-constrained environments. |
| Keywords: Lane detection Lightweight CLRNet MobileNetV3-Large Coordinate attention |
