The researchers, led by Weidong Li, Huiyong Xie, Zhenying Li, Daijun Zhang and Shuai Wang, call their approach the Mining Industry Unsafe Behaviors Monitoring Framework, or MIUBM-Net. The framework splits the detection problem into two complementary tasks. The first module, named STFNet-A, handles typical unsafe behaviors, the readily identifiable actions that violate safety protocols. The second, Alpha-SDL-Pose, focuses on posture-related unsafe behavior, where the danger lies not in what a worker is doing outright but in how the body is positioned in a mining environment.

The technical centerpiece of the posture module is a mechanism the authors call Dual Non-Local Attention, or DNLA, which they integrated into an optimized version of the AlphaPose skeleton-estimation architecture. Applying this dual attention design pushed posture recognition accuracy to 94.9 percent while sustaining a processing speed of 49.4 frames per second, a combination the authors describe as suitable for real-time monitoring.

For the typical-behavior module, STFNet-A reached a mean average precision of 96.9 percent at 59.6 frames per second on the researchers’ benchmark. Together, the two modules form a pipeline that the study’s experimental results indicate can recognize both categories of unsafe behavior in underground mining environments. Beyond the models themselves, the team emphasizes that the comprehensive miner unsafe behavior dataset they compiled addresses a long-standing obstacle in the field: the scarcity of dedicated training data for mining-specific safety scenarios, which has limited how well general-purpose recognition systems transfer to these settings.

The authors position the work as a practical step toward intelligent mining safety monitoring, arguing that achieving both high accuracy and real-time speed under complex mining conditions has been the central challenge holding back deployment. As with any computer-vision study, the performance figures reflect the researchers’ own experimental evaluations, and real-world robustness across varied mine sites, lighting conditions and camera placements would need to be demonstrated in practice. The framework nonetheless offers mining operators a concrete architecture for automated behavior monitoring, built on a dedicated dataset compiled specifically for unsafe behavior recognition in mining environments.

Subject of Research: Optimizing AlphaPose with Dual Non-Local Attention for Real-Time Unsafe Behavior Detection in Mining Environments

Article Title: Optimizing AlphaPose with Dual Non-Local Attention for Real-Time Unsafe Behavior Detection in Mining Environments

Article References: Li, W., Xie, H., Li, Z., Zhang, D., & Wang, S. (2026). Optimizing AlphaPose with Dual Non-Local Attention for Real-Time Unsafe Behavior Detection in Mining Environments. Machine Learning, 115(11), Article 252. https://doi.org/10.1007/s10994-026-07185-w

Image Credits: AI Generated

DOI: 10.1007/s10994-026-07185-w

Keywords: Optimizing, AlphaPose, Dual, Non-Local, Attention, Real-Time, Unsafe, Behavior, Detection, Mining, Environments, scientific research