关键词: 防护用具, YOLOv7, 可变形卷积, BiFormer, 泛化问题

Abstract: In the field of on-site protective equipment detecting, the data of the model mostly uses non real construction site backgrounds and only detects single safety helmets and protective equipment, when applied to specific scenarios, false positives and missed detections are prone to occur. To solve the above problems, a detection method for protective equipment of workers on construction sites from the perspective of drones is proposed. While improving the accuracy of algorithm detection, the practical applicability and generalization of the model have also been effectively improved. Using complex construction scenes captured by drone aerial photography as experimental datasets, and then annotating and preprocessing the data. The deformable convolutional module is introduced into the backbone network of YOLOv7 algorithm, which automatically adapts to the morphological changes of the safety helmet and warning suits. And the BiFormer attention module is embedded in the SPPCSPC module to improve the detection performance of the model for small-scale targets. To further improve performance of the model in localization and robustness to samples, WIOU is introduced as the regression loss function in the final prediction stage. Comparing this experiment with other algorithms on the self built construction site scene dataset and public dataset, the detection accuracy has been improved to a certain extent. It effectively validates the advantages of this algorithm in detecting protective equipment in complex construction site scenes.

Key words: protective equipment, YOLOv7, deformable convolution, BiFormer, generalization