FPGA Deployable Fire Detection Model for Real-Time Video Surveillance Systems Using Convolutional Neural Networks

Unchecked Fires can cause massive disasters leading to loss of human and animal life, property, forests, and other areas of vegetation and causing serious environmental damage. Moreover, fire has a tendency to spread rapidly over a small duration of time. Hence early detection is a very crucial factor when handling fire disasters. Recent studies have proposed the use of Convolutional Neural Network (CNN) architectures for the purpose of fire detection. In deep learning, a CNN is a class of deep neural networks, most commonly applied to analyzing visual imagery. CNNs have the ability to automatically learn features from raw fire data which eliminates the tedious and time consuming task of feature engineering. However, popular CNNs (like Alexnet, GoogleNet, VGGNet etc) are computationally expensive and bulky in size. Our research aims at exploring other lightweight alternatives without compromising the performance. This would make the CNNs easily deployable on Field Programmable Gate Arrays(FPGAs) and other mobile devices with limited resources. Deploying the CNN model onto FPGAs would make it portable and hence, it can be conveniently used as an efficient stand-alone fire detection system. In our research we have trained MobileNet, MobileNetV2 and DenseNet like architectures on our dataset and compared their performance on several metrics. Based on our experiments, we have developed a system that successfully detects fire in real-time scenarios and raises an alarm.