Fire Detection Based on Improved HOG

Rules in color model is an important method to identify candidate fire region. If the candidate fire region is not correct, the features extracted in subsequent steps are wrong. In this paper, a combination of constraint rules based on RGB and YCbCr color model is proposed. Compared with the constraint rules based on the YCbCr color model of other references, the proposed method extracts less noise from the candidate fire region, whether it is a fire picture or a non-fire picture. The HOG feature can not only detect the edge part of the object well, but also quantify the distribution structure of pixels inside the object through the distribution of statistical gradient vector angles. Based on these two advantages, this paper uses the HOG feature for fire detection. Scanning the entire image with overlap, the HOG feature vector dimension of the image is too large. Time complexity of fire detection is higher. This paper improves it and combines multiple features to detect fire. The average detection rate for fire video is 96.1%, and the average detection rate for none-fire video reached 100%.