Lightweight Gated Parallel Fusion of CoordAtt and ASPP for Degradation-Robust Monocular Parking-Line Segmentation

Existing vision-based parking slot perception methods predominantly adopt Around View Monitor (AVM) or Bird’s Eye View (BEV) imagery and target intact markings. Robust segmentation of heavily worn parking lines from monocular rear-facing cameras, however, remains largely unexplored. This paper proposes a lightweight fully convolutional network for degradation-robust monocular parking-line segmentation. Coordinate Attention (CoordAtt) and Atrous Spatial Pyramid Pooling (ASPP) operate in parallel at the encoder bottleneck, and their outputs are fused through a learnable spatial-adaptive gate that assigns pixel-wise weights to each branch, eliminating the mutual interference inherent in serial fusion. On a self-constructed dataset of 2273 annotated monocular parking-line images, the proposed network achieves 64.03% IoU and 280.5 FPS with 7.65 M parameters. Ablation experiments confirm that serial fusion degrades directional precision, whereas parallel gated fusion preserves both directional cues and multi-scale context through pixel-level adaptive allocation. These results demonstrate improved segmentation accuracy and wear-completion robustness under severe marking degradation, providing an efficient perception solution for monocular camera-based automatic parking systems.