Application Research on SGDM-YOLOv7-tiny Algorithm for Smoking Detection

In response to the current challenges in low accuracy and poor timeliness of smoking detection in public places, which is mainly implemented using infrared, temperature measurement, and other technologies, this paper proposes improvements to the YOLOv7-tiny algorithm, presenting a real-time smoking detection algorithm based on Slim-Gather-and-Distribute Mechanism(SGDM-YOLOv7-tiny). Firstly, an enhancement is made upon the EMA attention mechanism, introducing the EMSA attention mechanism to preserve more crucial features of the target objects, directing the network’s attention to smaller targets and thereby improving detection accuracy. Secondly, in the feature fusion phase, the GSConv and Slim Neck concepts are incorporated to refine the Gather-and-Distribute process, collecting and integrating information from all levels, distributing it across different levels, enhancing the partial information fusion capability of the neck, improving the performance of deep convolutional neural networks, and simultaneously reducing computational complexity and network structure. Lastly, a loss function, NWD, more suitable for measuring small targets, is employed to calculate the loss. Experimental results indicate that the SGDM-YOLOv7-tiny algorithm achieves increases of $2.27 \%$ and $4.25 \%$ in accuracy and recall, respectively. The mAP @ 0.5 and mAP @ 0.5: 0.95 also show improvements of $2.83 \%$ and $3.6 \%$, effectively enhancing the algorithm’s detection accuracy. This achieves a more real-time and accurate smoking detection, meeting the requirements, and providing feasibility for deployment on edge terminal devices.