Image based tiny fire detection in a grid environment using human geometric constrains

Abstract

Smoking detection is critical for fire safety in the grid facility environment. The vision-based smoking detection algorithm plays a key role to prevent a fire in grid facility environment. However, traditional methods based on target detection remain challenging in the realistic scenario, such as that it can not correctly distinguish whether a tiny object is a cigarette. In this paper, based on human geometric constraints(HGC), we propose a method which combines action recognition and target detection. First, in order to remove irrelevant action frames, we extract human skeletal key points from relevant frames using HGC algorithm, and build a smoking action dataset. Second, we train a neural network with the dataset of smoking action and then combine the deep learning-based target detection method to refine the judgment of smoking action with the bounding box extracted by skeletal points. Comprehensive experimental results demonstrate that our proposed method removes about 90% interfering action frames, improves average precision of the network from 77% to 82% in our dataset.