A Fire and Environmental Object Detection Method Baesd on Jointly Optimized YOLOv7

Fire detection systems predominantly rely on hardware detectors, these detectors face significant challenges, including susceptibility to false alarms, limited detection range, and varying effectiveness across different fire types. Recent advancements in computer vision and deep learning have significantly improved fire detection capabilities, yet challenges remain in accurately detecting small and obscured objects. This study aims to enhance the YOLOv7 framework to overcome these challenges. First, we replace certain ELAN modules in the head and backbone networks with ODConv2d, which improves the model’s sensitivity to small objects by assigning different attention scalars to different channels. Additionally, a vision transformer with a bi-level routing attention mechanism is integrated into the concatenation layer of the ELAN-W module in the head network, enhancing the network’s ability to detect smoke by filtering out irrelevant background information. Finally, the inclusion of environmental markers in fire scene images allows for the simultaneous detection of fire scenes and types, offering more comprehensive support for firefighting strategies. The proposed model achieves 99.6% in smoke detection accuracy, and a 5.8% improvement compared to the baseline, thereby providing a more robust solution for real-time fire detection in diverse environments.