Flame recognition and location algorithm based on YOLO v5s and binocular vision

Abstract An improved YOLO v5s framework in conjunction with binocular vision is developed to fulfill the stringent requirements for precise identification and accurate localization of flames. An Efficient Multi-scale Attention (EMA) mechanism is seamlessly integrated into the backbone network of the YOLO v5s model, thereby significantly improving the capacity to focus on and interpret critical target features. To bolster the multi-scale feature fusion capability, the Bidirectional Feature Pyramid Network (BiFPN) is strategically introduced at the fusion layer within the object detection architecture. The Semi-Global Block Matching (SGBM) algorithm is deployed to synchronize the binocular images, complemented by the least squares curve fitting method to adjust location inaccuracies, enhancing the precision of flame positioning. A binocular vision system is meticulously constructed to facilitate a series of rigorous experiments on flame detection and location. The experimental results demonstrate that the improved YOLO v5s network model achieves an exceptional flame recognition rate of 96.25%, which represents a noteworthy increase of 0.98% in accuracy in comparison to the original YOLO v5s model. Additionally, within a distance range of 3 to 10 meters, the flame ranging error is consistently less than 0.1 meters, successfully realizing the objective of high-precision flame detection and localization.