Enabling efficient WiFi-based occupant behavior recognition using insufficient samples

Abstract

To launch energy-efficient and occupant-centric services in smart buildings, understanding occupant behavior is becoming an essential underpinning. Prevailing approaches mainly adopt ambient sensors, wearables or vision cameras, which suffer from insensitive perception, arduous deployment or privacy-leakage risks. Recently developed WiFi-based methods circumvent the above limitations, however, most of them require sufficient training samples and computational resources for a comprehensive understanding of behavior patterns. This scenario requires occupant cooperation and considerable energy consumption. In this paper, we propose the use of LT-WiOB, which is a lightweight triplet framework for WiFi-based occupant behavior recognition, that provides a cost-efficient and user-friendly solution. Technically, the novelties of LT-WiOB are threefold. For the insufficient data problem, an efficient triplet framework is presented to measure the embedding dependencies; it attempts to avoid the ambiguity of new samples from each class and allows only a few samples to be entered. Regarding inefficient computation, a lightweight convolutional module is specifically designed and implemented to reduce the model complexity. For unstable training, the triplet architecture is further enhanced by employing a novel triplet sampling strategy that creates useful triplet candidates. Extensive experiments were conducted for performance comparison under diverse environmental settings. The results showed that LT-WiOB realizes the best overall accuracy of 96.1% and achieves over 90% and 85% accuracy in the five-shot and one-shot conditions, respectively, with almost half of the computation cost being saved. These results verify the efficiency and feasibility of the proposed LT-WiOB model.