Pedestrian Detection in Underground Coal Mines with an Improved YOLOv7 Algorithm

To achieve intelligent monitoring and unmanned driving in coal mines, accurate pedestrian detection plays a crucial role. However, due to various complex factors such as uneven lighting, dense dust, cable interference, and obstacles underground in coal mines, traditional detection methods often struggle to accurately detect pedestrians. To address this challenge, an improved algorithm based on the YOLOv7 network is proposed. This algorithm introduces a deformable attention mechanism in the Backbone, allowing it to dynamically adjust the shape and size of regions to focus on important features of the image. Additionally, in the channel boosting structure (CBS), an activate or not (ACON) function is used after batch normalization (BN), which adaptively adjusts the activation output based on the input amplitude. This helps to better maintain gradient flow and promote more stable and efficient training. Specifically, the Wise IoU (WIoU) loss function is introduced into the model. This loss function provides a comprehensive approach that considers both spatial overlap and confidence scores, enabling the model to learn more robust and accurate predictions. Experimental results show that the [email protected] reaches 96.7%, and the precision rate reaches 97.2%. These values are 1.5% and 1.9% higher than those of YOLOv7, respectively.