Work

Thermal descriptor extraction through reciprocal residual attention learning for construction worker unsafe condition detection and localization

Year: 2026

Type: article

Abstract: Unsafe worker actions in elevated construction environments remain a leading cause of occupational accidents, particularly under low-visibility and complex environmental conditions where conventional RGB-based monitoring systems exhibit limited robustness. To address these challenges, this study proposes a novel Thermal Context-Guided Dual-path Reciprocal Network (TCDRNet) for intelligent detection of unsafe thermal actions. Initially, a novel Thermal Adaptive Gradient (TAG) preprocessing filter is introduced to enhance thermal image quality via local normalisation, gradient-based thermal boundary enhancement, adaptive thermal weighting, feature amplification, and edge-preserving smoothing. The generated TAG images are subsequently processed utilising the proposed TCDRNet, which incorporates a Cross Stage Partial Dual-path Reciprocal Network (CSPDRNet) backbone integrated with Thermal Global Descriptor Extraction (TGDE), Reciprocal Attention Generation (RAG), and Residual Reciprocal Fusion (RRF) modules to facilitate reciprocal thermal feature learning. Furthermore, Spatial Pyramid Pooling (SPP) is employed for multi-scale contextual extraction. Concurrently, the Path Aggregation Network (PANet) performs hierarchical feature aggregation to generate low-level localisation features, mid-level boundary information, and high-level semantic representations. Finally, a Thermal Context-Aware Residual Attention (TAR) module is introduced to refine the multi-scale features via channel attention, spatial attention, and residual fusion, thereby emphasising discriminative unsafe regions. Experimental evaluation on a thermal unsafe action dataset comprising no-helmet, no-shoes, mobile-phone utilisation, safe and unsafe posture classes demonstrates the effectiveness of the proposed framework. TCDRNet achieves an accuracy of 96.60%, precision of 96.41%, and recall of 95.96%, outperforming several state-of-the-art models.

Source: Scientific Reports

DOI: 10.1038/s41598-026-76021-x

Authors: Jeong In Choi, Shyamala Devi M, K. R. Sri Preethaa, Yuvaraj Natarajan, Yirong Mi

Institutions: Soonchunhyang University, Soonchunhyang University Hospital Bucheon, Kyungpook National University, Vellore Institute of Technology University

Language: English

Topic: Human Pose and Action Recognition

Subfield: Computer Vision and Pattern Recognition

Field: Computer Science

Domain: Physical Sciences

SDGs: Good health and well-being, Industry, innovation and infrastructure

Open Access status: gold

Estimated APC paid: $2,690