Improved YOLOv8-based method for detecting non-compliant behaviors of workers in smart factories

In response to the challenging issues of detecting non-compliant behaviors of workers in smart factories, such as insufficient real-time performance, low accuracy, and occlusions caused by complex factory environments, a method based on an improved YOLOv8 model for detecting non-compliant behaviors of workers is proposed. This method introduces dynamic snake convolution(DSConv) into the backbone network of YOLOv8 to enhance the capturing capability of detailed features and improve the overall feature extraction ability of the network. Additionally, the efficient multi-scale attention (EMA) mechanism is embedded into the backbone network to alleviate the influence of occlusions and background interference, allowing the model to focus on body parts or actions related to non-compliant behaviors. To further enhance the training efficiency and regression accuracy of the model, the Alpha generalized intersection over union (Alpha_GIoU) loss function is employed to optimize the complete intersection over union(CIoU) loss function component of the YOLOv8 model. This takes into account the aspect ratio and position information of the target box, accelerating the convergence speed of the model. Experimental results demonstrate that the improved YOLOv8 achieves a mAP@0.5 of 94.98%, significantly outperforming the baseline YOLOv8 model and other mainstream methods. Moreover, the detection speed reaches 53.8 FPS, meeting the requirements of high efficiency and accuracy for detecting non-compliant behaviors of workers in smart factories.