• deep learning
• embedded systems
• image processing
• neural network

Design and analysis of a DNN architecture for smart fire and smoke detection in indoor/outdoor environment. The algorithm uses a two-subnetwork architecture joined with an alarm-triggering code, it was trained on custom dataset and tested versus most recent state of the art solutions and traditional AI-based methods through the computation of performance indices such as Accuracy, Precison, Recal and F1-score. The algorithm deployment on low-end embedded systmes, like Raspberry Pi 3 and Nvidia Jetson Nano, is then discussed in detail confronting the resulting performance against both deep-Learning-based and non-DL-based solutions from the State of the Art and traditionally employed techniques such as fine-tuned VGG DNN. Results show that the proposed solution outperform benchmarks in term of performance indices while achieving computational cost comparable to non-DL-based solutions.