Seat Belt Wearing Detection Based on EfficientDet_Ad

Seat Belt Wearing Detection Based on EfficientDet_Ad

연구 분야: Verification

학회: International Conference on Computer Animation and Social Agents

For the existing safety belt target detection algorithm in the realm of electrical power scenarios, which faces challenges such as low precision in recognizing small targets, easy confusion of target features with the background, and limited computational resources on the intelligent monitoring platform. In this paper, we propose an EfficientDet-based safety belt detection algorithm, EfficientDet_Ad. Firstly, an enhanced feature extraction module is designed for multi-scale feature fusion, improving the capability to detect small targets. Secondly, a feature fusion attention module, FFAM, is constructed to enhance the focus on less prominent targets. Finally, the algorithm achieves a reduction in network parameters by introducing Ghost Conv and Channel Shuffle operations to reconstruct the MBConv module in the EfficientNet backbone. Analyzing the results of the comparative experiment and ablation study reveals that the proposed EfficientDet_Ad algorithm achieves an average precision of 90.12%, surpassing the EfficientDet algorithm by 6.64%, and outperforming other advanced object detection algorithms. Simultaneously, the detection speed reaches 55.7 frames per second. The comprehensive experimental results demonstrate the remarkable effectiveness of the algorithm in terms of accuracy and real-time performance.