A Lightweight Model for Human-Object Interaction Recognition in Underground Coal Mines Based on Instance Segmentation

Existing behavior recognition methods for underground coal mines(UCM) are often limited to single scenarios, typically focusing on detecting and warning of specific abnormal behaviors, such as the hazardous action of crossing conveyor belts. However, miners working underground for extended periods engage in a variety of actions and operations, which are closely linked to the environment and the objects they interact with. Fine-grained recognition of these behaviors is essential for enhancing intelligent monitoring capabilities and improving safety in coal production. To address this, we propose a lightweight recognition model for human-object interaction (HOI) behavior in UCM based on instance segmentation. The model simultaneously monitors miners and relevant objects, using an instance segmentation network to capture the location and shape information of both humans and objects. By integrating this information and applying rule-based reasoning, the model achieves accurate and fine-grained recognition of miners' behaviors. Experimental results show that the proposed model reaches a new level of accuracy in detecting miners, objects, and behavior recognition, with model parameters of only 2.3M and an inference speed of 120 fps, demonstrating superior efficiency and scalability. This model significantly enhances intelligent monitoring in UCM and improves overall safety and productivity.