Recognize Locomotion and Transportation Modes from Wi-Fi Traces via Lightweight Models

With the ever-increasing intelligent perception capability of mobile terminals, the demand for fine-grained locomotion and transportation mode recognition (LTMR), a special branch of human activity recognition, has become imperative. However, existing works that utilize GPS information for LTMR raise practical concerns regarding spotty coverage and battery depletion, while the inertial measurement unit (IMU) databased model's generalizability has also been questioned. Towards this end, we propose recognizing locomotion and transportation mode from Wi-Fi traces via lightweight models motivated by the accessibility of Wi-Fi information and the ubiquitousness of Wi-Fi access points (APs). After data pre-processing, we divide time windows to construct the samples, acquire statistical features, and supply these features to various lightweight models. We evaluate our proposed methods on the Sussex-Huawei Locomotion-Transportation (SHL) recognition challenge dataset, and the results demonstrate that our methods achieve a maximum 90.08% macro-F1 score on the test dataset, and LTMR from Wi-Fi data can be performed with minimal memory footprint and computational resources.