Lightweight Parking-Line Segmentation for Automated Parking in Underground Garages via Shallow Detail Injection and Deep Semantic Slimming

Parking-line segmentation for automated parking in underground garages is challenged by sparse targets, perspective-compressed thin structures, and visually similar disturbances such as cracks, joints, uneven illumination, and reflections. We propose a lightweight YOLO26n-sem framework with Semantic-Preserving Shallow Residual Detail Injection (SRDI) and Interface-Preserving Deep Semantic Slimming (IPDS). SRDI injects scale-aligned P2 details into the P3 pathway through a bounded learnable residual coefficient, while IPDS reduces hidden widths in deep P5 modules without changing the external P5 interface or top-down propagation topology. On a four-class bird’s-eye-view dataset acquired by a vehicle-mounted surround-view system, the model achieves 85.966% mIoU, 83.081% Parking IoU, and 87.700% Thin Recall, improving YOLO26n-sem by 0.500, 0.973, and 1.556 percentage points, respectively. Parameters decrease by 24.9%, from 1.631 M to 1.225 M, and latency decreases from 8.8203 to 7.0321 ms (142.21 FPS). Zero-shot evaluation on an independent public underground-garage dataset achieves 77.837% mIoU and 72.625% Parking IoU, exceeding the baseline by 1.416 and 2.917 percentage points, respectively, without target-domain adaptation. Overall, the proposed model improves segmentation accuracy while reducing parameter count and inference latency, and its advantage over the baseline remains observable under the tested cross-garage domain shift.