LabSeg: Semantic Segmentation for Aisle Obstruction and Bench Clutter Monitoring

Detecting aisle obstructions and assessing bench clutter in laboratory environments are critical for ensuring safety and operational efficiency in scientific research. However, existing semantic segmentation methods struggle in such settings due to severe occlusion under fixed viewpoints, scarcity and homogeneity of training data, and the unique challenges of nighttime infrared imaging. To address these issues, we propose LabSeg, a tailored semantic segmentation framework built upon YOLO11-seg, with three key components: (1) To capture long-range spatial dependencies essential for aisle obstruction detection, we extend the C2PSA attention mechanism with spatial coordinates, yielding the SC-C2PSA module; (2) To better model the spatial distribution of clutter relative to static laboratory infrastructure, we introduce a Context Enhancement Module (CEM) before the segmentation head to refine foreground-background contextual understanding; (3) To improve robustness under varying illumination, especially in nighttime infrared conditions, we design a Lighting-Invariant Preprocessing (LIP) layer that normalizes cross-spectral input characteristics. Experiments on our self-collected LabSafety-2025 dataset show that the proposed method improves aisle obstruction detection accuracy by 8.1% and increases the F1-score for bench clutter assessment by 10.3%. On the COCO val2017 benchmark, the medium-scale variant (YOLO11-seg-m) without LIP improves Mask mAP50 by 1.1%, demonstrating its generalization capability. This work provides a practical solution for intelligent laboratory monitoring and contributes methodological insights for semantic segmentation in constrained, low-data, and multi-spectral environments.