Development of an Early Fire Detection Technique Using a Passive Infrared Sensor and Deep Neural Networks

Early detection of fire is key to mitigate fire related damages. This paper presents a differential pyro-electric infrared (PIR) sensor and deep neural networks (DNNs) based method to detect fire in real-time. Since the PIR sensor is sensitive to sudden body motions and emits a continuous time-varying signal, experiments are carried out to collect human and fire motions using a PIR sensor. These signals are processed using one-dimensional continuous wavelet transform to perform feature extraction. The corresponding wavelet coefficients are converted into RGB spectrum images that are then used as inputs for a deep convolutional neural network. Various pre-trained DNN architectures are adopted to train and identify the collected data for background (no motion), human motion, and fire categories: small quasi-static and spreading fires. Experimental results show that the ShuffleNet architecture yields the highest prediction accuracy of 87.8%. Experimental results for the real-time strategy which works at a speed of 12 frames-per-second show 95.34% and 92.39% fire and human motion detection accuracy levels respectively.