Abstract:

A training-free collaborative perception framework named MineGuard was proposed to address the domain generalization challenge of unsafe personnel behavior detection in complex and dynamic coal mine scenarios, as well as the industrial pain points of scarce annotated data and high adaptation costs for new scenarios. Firstly, the general cognitive capabilities of vision-language large models (VLMs) were integrated with the specialized detection capabilities of lightweight expert models to build a cognition-execution layered architecture. Then, a semantics-driven collaborative reasoning mechanism was established to enable intelligent scheduling and collaborative verification of multiple models. Finally, a cross-mine evaluation benchmark was constructed to validate the framework's generalization performance. The results show that on the coal mine unsafe behavior dataset, MineGuard reaches the mean average precision (mAP) of 90.22% without any fine-tuning, which is 5.65% higher than traditional source-domain-trained methods. It exhibits excellent cross-domain generalization performance.

Key words: coal mine, unsafe behavior, training-free, domain generalization, large model