An End-to-end Campus Fire Detection Algorithm Incorporating Priori Features

To address the imbalance in detection speed and accuracy caused by smoke, flame, and background confusion in campus video surveillance scenarios, an end-to-end campus fire detection algorithm incorporating priori features is proposed. First, a feature pyramid fusion module guided by priori features is constructed to cascade the multi-scale feature fusion module. The priori features are individually pre-trained by Convolutional Neural Network (CNN) classification network. The feature complementarity between the two is utilized to improve the feature representation of the detection network. Second, a lightweight set prediction detection head is designed. The head does not require manually designed anchors and post-processing strategies, enabling end-to-end detection while balancing speed and accuracy. Finally, a joint loss function that integrates the losses of categorical and bipartite matching is proposed. Focus on the drawbacks of the slow convergence associated with bipartite matching, both classification information and bipartite bit matching are considered in the detection results. Experimental results on the fire detection dataset demonstrate that the proposed algorithm improved average accuracy by 7.8% while maintaining detection speed, effectively balancing speed and accuracy in fire detection.