Embedded Files
Lane Detection Overview
The first three types of lane detection methods shown in the figure represent the mainstream approaches currently used in lane detection systems. Each of them employs a task-specific module designed to handle either lane existence classification or lane coordinate prediction, depending on the architecture.
(a) Segmentation-based lane detection methods (e.g., SCNN, RESA) rely on dense pixel-wise mask prediction to detect lane areas, followed by post-processing to extract individual lane curves.
(b) Anchor-based lane detection approaches (e.g., LaneATT, ADNet) introduce predefined anchors and refine their positions and classifications to predict lanes.
(c) Parameter-based lane detection (e.g., LSTR) directly regresses the shape parameters of lane curves and their spatial offsets, enabling efficient and interpretable predictions.
Despite architectural differences, all three types adopt task-specific head networks that specialize in addressing the unique requirements of lane detection, such as handling occlusions, maintaining geometric consistency, and ensuring high runtime performance.
Backdoor attack strategy-specific heatmap
Building on these foundations, we design a unified gradient-based approach to identify optimal trigger positions for backdoor injection. For a given input image and selected attack strategy—namely Lane Disappearance Attack (LDA), Lane Offset Attack (LOA), or Lane Reconstruction Attack (LRA)—we first determine the appropriate loss function according to the model type and task formulation (as summarized in Table 1). Specifically, classification losses such as focal loss or cross-entropy (CE) are used for LDA to suppress lane existence predictions, while regression-oriented losses like L1, BCE, or GLIoU are employed in LOA and LRA to induce geometric distortions. By computing the gradient of the selected loss with respect to the input, we obtain a task-specific heatmap that highlights
sensitive regions most influential to the model’s output.
Page updated
Google Sites
Report abuse