Toward Reliable Fire Detection in Indoor CCTV Footage: Reducing False Alarms From Benign Flames Using 3D Attention CNNs

Reliable early fire detection is essential for protecting life and property. While deep learning–based fire detection systems have demonstrated potential, many remain untested under real-world conditions, particularly in indoor environments where false alarms are prevalent. Existing studies primarily address limited false-positive cases—such as bright illumination or red-colored objects—while neglecting realistic, benign flame sources including vented fireplaces, gas fire pits, candles, and LED flame simulators. To address this gap, we introduce FalseNet, the first curated video dataset encompassing a diverse set of non-hazardous indoor flame scenarios alongside genuine fire incidents. We also propose a 3D attention-based convolutional neural network that integrates ResNet with (2+1)D convolutions—where a 3D convolution is factorized into sequential 2D spatial and 1D temporal operations—and Convolutional Block Attention Modules (CBAM) to enhance spatiotemporal feature extraction. The proposed model achieves an accuracy of 90.44%, precision of 90.46%, and recall of 90.44%, demonstrating improved robustness in reducing false-positive rates. Compared to a baseline model without CBAM, the results confirm a notable performance gain, validating the role of attention mechanisms in enhancing model generalization. Grad-CAM visualizations illustrate the model’s ability to localize spatial features associated with benign flames, highlighting its capacity to differentiate benign fire instances from actual fire events. These findings support the viability of the proposed approach for intelligent, CCTV-based fire detection with enhanced reliability in complex indoor environments.