| Popis: |
To solve the problem that the current fire detection algorithm is susceptible to false positives and false negatives due to the influence of complex scenes and lighting conditions, this paper proposes a video fire recognition algorithm based on attention mechanism. Firstly, video stream datum are read frame by frame, then used for dynamic and static judgment. Fire features are extracted by static color segmentation and dynamic flicker judgment in the YCbCr color model. Secondly, multi-feature fusion and attention mechanism are introduced into the ResNet-50 network, which is used to automatically extract fire features to improve the detection effect of small targets. Finally, center point detection is used to predict fire and realize rapid and accurate fire positioning. The experimental results show that the processing speed of the algorithm is 68FPS and mAP is 89.1% on the fire dataset established in this paper. It has good robustness and meets the needs of real-time video fire detection. |
