Promoting Occupancy Detection Accuracy Using On-Device Lifelong Learning

Our recently developed synchronized low-energy electronically chopped passive infrared (SLEEPIR) sensor node enables the stationary occupancy detection capability of traditional passive infrared (PIR) sensors. A machine learning (ML) algorithm reports occupancy based on a locally collected dataset from the sensor node. Though promising, the ML algorithm’s detection accuracy depends on the diversity of the collected dataset—provided that the dataset contains a wide variety of infrared (IR) noise and occupancy patterns. Thus, it is challenging to train a universal ML model that contains all possible patterns. We propose an efficient ${K}$ -nearest neighbor (KNN) occupancy classifier that incrementally adapts to the novel data from the sensor. The proposed algorithm ensures that only the relevant noise and occupancy patterns are learned. The fact that training observations are gathered on the same sensor node where the inference is made keeps the proposed classifier accurate even with the bounded size of the dataset. A small dataset and an architecture like KNN both enable the training and inference to be executed on a resource-constrained Internet of Things (IoT) device. Thus, the proposed on-device lifelong learning (ODLL) approach eliminates the need for over-the-cloud ML model updates. The dataset was collected for two distinct floorplans over two months. Results indicate an average occupancy accuracy improvement of 20.8% compared to a statically trained long short-term memory (LSTM) model. The proposed KNN model delivers comparable detection accuracy while remaining orders of magnitude faster in terms of computational performance when compared to the LSTM-based occupancy detection algorithm.