You have no items in your shopping cart.
ABSTRACT
Fire outbreaks are events that pose a huge threat to lives and property, as such, it is of utmost importance that detecting fire from scenes require a level of accuracy and speed that matches its importance. Most existing methods to detect fire from scenes use sensors to capture data like smoke ionization, heat signatures. However, these methods are prone to errors from ambient heat, and distance from the flame to the sensor which increase detection latency especially in well-spaced and high-ceilinged areas.
In addressing these challenges, this work proposes a novel approach to flame recognition and localization using a deep learning model deployed on IoT surveillance systems. The deep learning model is a lightweight architecture that performs inference on the cloud at low latency and, it was developed using the pre-trained SqueezeNet convolutional base and a custom trained two-layer neural network classifier. This model is deployed to Google-cloud’s AI Platform prediction service for performing fire inference on streaming video data. The IoT surveillance system consists of a network of low-end cameras designed to stream video data to a centralized local-server that parses these frames with image processing techniques. This local-server is developed to upload the frames to specific folders unique to the cameras on the cloud bucket. On finalizing each upload, a developed cloud function called the predictor is triggered that invokes the model through an API call for predictions. The predictor receives the predictions and sends it to another developed cloud function called the publisher that parses data from the predictor and publishes the prediction to all authorized subscriber applications using the Google Cloud Pub/Sub service
The system achieves a prediction latency of 2s per frame and SqueezeNet a prediction accuracy of 89% on a test set gathered from public data repositories.
