Structure-Enhanced Features and Quality-Aware Dynamic Anchor Scoring for Robust Lane Detection

The paper proposes a structure-enhanced and quality-aware framework for robust lane detection, built upon the Anchor Decomposition Network (ADNet). The framework addresses two coupled failure modes in anchor-based lane detectors: backbone features losing structural continuity along partially visible lanes, and classification confidence decoupling from line-level localization quality, which allows inaccurate anchors to persist before non-maximum suppression (NMS).

The first proposed component is the Gated Horizontal-Vertical Token (GHVT) module, a lightweight residual backbone plug-in applied to ResNet stages 3 and 4. It models directional token interactions along horizontal and vertical axes and injects the resulting context through a learnable residual gate initialized to 10−4. The module is detector-agnostic, preserving feature tensor shapes and leaving the detector head, geometry decoder, and NMS pipeline unchanged.

The second component is Line-Quality-Aware Dynamic Anchor Scoring (LQAS), which recalibrates ADNet's existing classification logits so pre-NMS scores better reflect line-level localization quality. LQAS combines quality supervision using LineIoU-based targets, hard-negative suppression, and pairwise ranking, without adding an inference branch. During training, only the classification layer is updated while the backbone, GHVT, FPN, SPGU, ALAU, and regression head are frozen.

On the VIL-100 benchmark, the enhanced ADNet-R34 model improves the reproduced baseline from 89.97 to 91.28 in F1@50, reducing both false positives (from 4.43% to 4.26%) and false negatives (from 4.90% to 4.69%). It also outperforms the reported ADNet-R101 results despite using the smaller ResNet-34 backbone.

On the full 34,680-frame CULane test split, the enhanced ADNet-R34 improves from 78.95 to 79.09 in F1@50 and from 60.65 to 60.86 in F1@75. The enhanced LaneATT-R34 improves from 76.58 to 76.75 in F1@50 and from 53.62 to 54.63 in F1@75, demonstrating cross-detector transfer. On TuSimple, F1@50 improves from 97.31 to 97.49 while the FP rate drops from 2.83% to 2.41%.

Ablation studies show GHVT provides the dominant gain, improving F1@50 from 89.97 to 91.21 (+1.24), primarily by reducing false negatives and increasing true positives. Adding LQAS further improves F1@50 to 91.28, mainly by reducing false positives from 738 to 712 while maintaining similar recall. Subset analysis shows GHVT gains are especially clear in occlusion-related subsets, with F1@50 increases of 1.68, 1.28, and 1.48 points for partial, severe partial, and full occlusion subsets, respectively.

Score-distribution diagnostics show LQAS reduces the hard-negative 90th-percentile score from 0.262 to 0.238 and drops the proportion of hard negatives above 0.30 from 8.73% to 8.02%, while preserving positive scores. Runtime analysis shows the enhanced ADNet-R34 increases parameters by 1.32% and reduces FPS by 3.31%, while LaneATT-R34 increases parameters by 1.31% and reduces FPS by 2.16%, indicating minimal computational overhead.

Improvements to AI systems:

1. Add a plug-in module for directional context modeling – Integrate the Gated Horizontal-Vertical Token (GHVT) module into any CNN backbone (e.g., ResNet stages 3/4) to preserve structural continuity along partially visible objects. This is detector-agnostic, adds 1.3% parameters, and reduces false negatives in occlusion-heavy scenarios.

2. Replace classification confidence with quality-aware scoring – Implement Line-Quality-Aware Dynamic Anchor Scoring (LQAS) to recalibrate pre-NMS scores using LineIoU-based targets, hard-negative suppression, and pairwise ranking. This decouples classification from localization quality, reducing false positives without adding inference branches.

3. Freeze-then-fine-tune training strategy – During LQAS training, freeze backbone, feature pyramid, and regression heads while updating only the classification layer. This allows targeted quality recalibration with minimal computational cost and prevents feature drift.

4. Cross-detector transferable scoring – Apply the LQAS recalibration to different anchor-based detectors (e.g., LaneATT) to improve F1@75 scores significantly (e.g., +1.01 on CULane) while keeping inference speed nearly unchanged.

What the improved AI system can do:

• Detect partially or fully occluded lanes (or analogous linear structures like roads, cables, or medical vessels) with higher recall, especially under occlusion (F1@50 gains of 1.28–1.68 points on occlusion subsets).

• Reduce false positives by suppressing low-quality anchors before NMS, improving precision without sacrificing recall (FP rate drops from 4.43% to 4.26% on VIL-100).

• Achieve higher F1 scores at stricter IoU thresholds (e.g., F1@75 improves from 60.65 to 60.86 on CULane), enabling more precise localization for downstream planning or mapping.

• Maintain real-time performance (FPS drop < 3.5%) while adding robustness, making it suitable for embedded or autonomous driving systems.

• Transfer the quality-aware scoring mechanism to other detectors with minimal retraining, improving their accuracy without architectural changes.

Lane detection requires recovering thin, elongated, and frequently occluded lane structures under challenging driving conditions. While anchor-based detectors provide efficient candidate generation, their performance is limited by two coupled issues: backbone features often lose structural continuity along partially visible lanes, and classification confidence may decouple from line-level localization quality, allowing inaccurate anchors to persist before non-maximum suppression (NMS). We propose a structure-enhanced and quality-aware framework that improves lane representation and dynamic-anchor scoring while preserving the inference pipeline of the Anchor Decomposition Network (ADNet). Specifically, a Gated Horizontal-Vertical Token (GHVT) module enhances mid- and high-level backbone features via lightweight directional token interactions with a learnable residual gate. In parallel, Line-Quality-Aware Dynamic Anchor Scoring (LQAS) calibrates existing classification logits using quality supervision, hard-negative suppression, and pairwise ranking without adding inference branches. On the VIL-100 dataset, our method improves ADNet-R34 from 89.97 to 91.28 in F1 score at the 0.5 intersection-over-union threshold (F1@50), reducing both false positives and false negatives. Additional experiments on CULane and TuSimple datasets, extensive ablations, score-distribution diagnostics, and runtime analysis confirm complementary structural and ranking improvements with minimal computational overhead.