Abstract:

To address the issue of insufficient small-object detection accuracy in remote monitoring of heavy industrial workshops, an unsafe behavior detection algorithm based on improved YOLOv7 was proposed. First, the traditional upsampling was replaced with a lightweight content-aware reassembly of features (CARAFE) module, which effectively preserved the semantic information of small objects through adaptive feature reassembly. Second, an improved Bi-level routing efficient layer aggregation network(Bi-ELAN) module was proposed by integrating the BiFormer dynamic sparse attention mechanism into the head network, which strengthened the multi-scale feature fusion capabilities and established target-background contextual relationships. Third, the loss function was refined by introducing the shape intersection over union(ShapeIoU)loss function, which enhanced bounding box regression accuracy through geometric shape constraints. Finally, ablation experiments and comparative experiments were conducted on the improved YOLOv7 model based on constructed remote monitoring perspective dataset. The results show that, while maintaining model lightweight characteristics, the proposed algorithm significantly improves small-object detection accuracy in remote monitoring scenarios. The improved model achieves a precision of 84.2%, a recall of 78.6%, and a mean average precision (mAP@0.5) of 78.8%. Compared to the original YOLOv7 algorithm, the improved algorithm increases precision, recall, and mAP@0.5 by 5%, 0.3%, and 2.6%, respectively.

Key words: remote monitoring, heavy industrial workshop, unsafe behavior, object detection, YOLOv7