Video smoke detection method based on cell root–branch structure

In recent years, smoke detection methods based on deep learning have become a research focus for the timely and accurate detection of industrial fires. However, in practical engineering applications, the detection rate of deep learning methods is lower when light-colored smoke with high transmittance blends with the background and foreground. To address this issue, we propose a video smoke detection method based on the cell root–branch structure, integrating various smoke characteristics. Firstly, suspicious target areas are selected using foreground object extraction methods and HSV color thresholding, focusing on the characteristics of smoke movement and whiteness. Then, for the clustering characteristic of smoke, we construct a cell root–branch structure model of the smoke grid. Areas with the same root are treated as a single smoke cluster, and large smoke clusters are selected based on the number of nodes. Finally, to differentiate from other potential targets, we use spatiotemporal context information and diffusion detection algorithms in video sequences, focusing on the characteristic of smoke diffusion, to determine suspected smoke areas. Compared with current mainstream object detection algorithms like Fast R-CNN and YOLOv8, our proposed method shows higher detection rates and lower false alarm rates in environments with complex backgrounds and multiple interference factors, validating the feasibility and robustness of our algorithm.