An Integrated Smoke Detection Method based on Convolutional Neural Network and Image Processing

Fire is one of the disasters against the Safety of human life and property. Generally, early smoke features are more obvious than fire in the surveillance environment. However, due to the variability of smoke characteristics (e.g., color, shape) and the interference of smoke-like objects (e.g., clouds, rivulet, and fog), the main challenge of smoke detection is false alarms in real-world. To tackle this problem, an integrated method is proposed which combines HSV color space, background subtraction with Faster R-CNN. This method can enhance smoke feature, meanwhile, it can reduce disturbance from smoke-like objects. Furthermore, a dataset is created which are collected from surveillance cameras that are installed in the wild. Our experiments show that the integrated method is more accurate and robust than previous work.