Structured lane instance detection with instance activation query and adaptive cross-attention

Abstract

Lane detection is a core task in the autonomous driving perception system, and the reliability of path planning and control can be directly determined by its performance. Existing methods are limited by rigid modeling paradigms or insufficient feature representation ability, making it difficult to adapt to the detection requirements of lane lines of arbitrary shapes in complex scenarios. To address this, this paper proposes an arbitrary-shaped lane detection method that integrates segmentation-aided supervision and structured modeling. The core innovations of this method are as follows: designing an Instance Lane Activation-based Query Generation module, which integrates Instance Lane Activation Map and Instance Lane Feature Aggregation to accurately generate instance queries focusing on effective lane line regions; introducing a spatially adaptive cross-attention mechanism to enhance the feature representation of slender targets through reference point-guided adaptive feature sampling; adopting a shape-unconstrained structured modeling paradigm to represent lane lines as arbitrary point sets to break through traditional shape limitations. Experiments on a self-constructed complex scenario dataset show that the proposed method outperforms mainstream methods in terms of detection accuracy and robustness under complex working conditions such as sharp turns, inclinations, and intersections, and can effectively achieve end-to-end accurate detection of lane lines of arbitrary shapes.