Pet immunity for PIR sensors using deep learning

Despite the ubiquity and wide spread success of deep learning for various tasks ranging from image classification to text generation, the adoption in the world of embedded systems and IoT seems relatively slower. In our work, we investigate possible application of this area of research to tackle the problem of home surveillance, namely pet immunity of passive infrared sensors (PIR). The classification whether the sensor was triggered by a person or a domestic pet is based on sequence of images taken by a camera attached to the sensor. The main challenge are extremely restricted resources of the on-board ARM CPU, especially memory in the order of kilobytes. We propose a simple approach based on background subtraction and classification using convolutional neural network. We demonstrate how despite its simplicity, it achieves scores that are competitive with much larger models employing pretrained open source detectors.