An Improved YOLOv10 for Helmet Detection in Complex Scenarios

Small target detection in complex backgrounds has always been a challenging problem to solve. For example, in construction sites, traditional methods for monitoring hard hat wearing to ensure personnel safety suffer from high labor costs and inefficiency. With the continuous development of object detection, significant improvements have been made in terms of accuracy and real-time performance. However, when it comes to detecting targets with single-feature characteristics like hard hats, current methods still fall short of achieving satisfactory results. In this study, we propose the BE-YOLOv10 network model for safety helmet detection. This model employs a BiFPN structure to alter the conventional feature extraction method, enabling the capture of richer feature and location information during the learning process. Additionally, inspired by the rank guidance introduced in v10, we incorporate the EMA attention mechanism into the C2fCIB stage. By combining this with the rank guidance of CIB, we redistribute weights and prioritize relevant features to enhance feature extraction and reduce computational redundancy, thereby improving model efficiency. Our algorithm achieves mAP of 65.3% on a self-constructed dataset, surpassing the previous model by 4.3%, with a reasoning speed of 0.4ms. These results fully demonstrate the superiority of our detection model algorithm in detecting small targets such as safety helmets.