关键词: 车道线检测, 深度学习, 多尺度空间注意力机制, 自适应特征融合

Abstract:

Aiming at the problem of lane detection accuracy of intelligent vehicles in complex scenes, this paper proposed a lane line detection algorithm which incorporates a multi-scale spatial attention mechanism and a path aggregation network (PANet). The algorithm first introduced the pre-anchored frame UFLD lane detection model and incorporated a feature pyramid enhancement module PANet with depthwise separable convolution to achieve multi-scale feature extraction of images. Next, a multi-scale spatial attention module was designed in the network framework and a SimAM lightweight attention mechanism was introduced to enhance the focusing ability on target features. Then, an adaptive feature fusion module was designed to perform cross-scale fusion of feature maps output from PANet by intelligently adjusting the fusion weights of feature maps at different scales, so as to effectively enhance the network’s ability to extract complex features. Finally, the application of TuSimple dataset detection proves that the proposed algorithm achieves a detection accuracy of 96.84%, representing a 1.02 percentage point improvement over the original algorithm, and outperforms conventional mainstream algorithms. Experimental results on the CULane dataset demonstrate that the proposed algorithm achieves an F₁ score of 72.74%, outperfor-ming conventional mainstream methods with a 4.34 percentage point improvement over the baseline. Notably, it exhibits significant performance gains in extreme scenarios (e.g., strong illumination and shadows), confirming its superior detection capability in complex environments. In addition, the real-time test shows that the model infe-rence speed reaches 118 f/s, which meets the real-time demand of intelligent vehicles.

Key words: lane line detection, deep learning, multi-scale spatial attention module, adaptively feature fusion