Zigzag Attention: A Structural Aware Module For Lane Detection

Lane detection presents a formidable challenge in the realm of autonomous driving, given the real-time processing demands, diverse acquisition conditions, and the unique elongated and angular characteristics of lane lines. While a multitude of network design strategies have been proposed to tackle this challenge, few effectively address the distinct morphology of lane lines. Approaches that leverage global relationships across all positions, e.g. attention mechanisms, are often hard to meet real-time processing requirements due to their computational complexity. To address these challenges and unique attributes of lane lines, including issues like occlusion and dashed lines, we present a specialized and plug-and-play attention module. It employs zigzag transformations to cohesively assemble spatially disparate, lane-relevant regions, thereby transforming the challenge into one of localized feature learning, which can be easily enhanced via lightweight convolutions and fully connected layers. Additionally, we harness the symmetry inherent in lane lines to bolster the learning process and enhance accuracy. Comprehensive experimentation validates the efficacy of our proposed module across a range of algorithms, demonstrating superior performance metrics, including parameters, computational complexity, and runtime, when compared to other attention approaches.