Improved Fire and Smoke Detection Method Based on RT-DETR

Fire and smoke detection is an important task for preventing early fires and reducing human casualties. Recognizing and classifying fire and smoke instances in various real-life scenarios based on infrared, smoke, and other sensors cannot achieve good accuracy. An improved flame detection algorithm based on RT-DETR is proposed to enhance fire and smoke detection accuracy. The BasicBlock in resnet18 is improved using the shift-wise operator, which changes the spatial location of the features, enabling the model to capture more spatial information. The CCFF in RTDETR is improved using HS-FPN, which enables the fusion of high-level features with low-level features. Finally, the improved RT-DETR model has an mAP value of 61.8% on the fire and smoke detection dataset of robo-flow, which is 1.7 percentage points higher than the original RT-DETR model, and also the number of parameters decreases by 29.0%, which improves the accuracy and achieves a lighter model at the same time.