Human Activity Recognition Scheme Based on Time-Series Analysis of Multi-Antenna AoA Data

This paper introduces an innovative indoor activity recognition system leveraging multi-antenna Bluetooth Angle of Arrival (AoA) data to address limitations in existing wearable, non-wearable, and signal-based recognition methods. Conventional systems often suffer from usability issues, limited location flexibility, or inadequate activity coverage. At the same time, signal-based methods such as WiFi Channel State Information (CSI) typically struggle with model adaptability across environments. Building on our previous single-antenna AoA approach, which captured azimuth and elevation angles but encountered difficulties in accurately detecting vertical movements and certain activities, this study proposes a dual-antenna solution for improved multidimensional sensing. Our system employs two AoA antenna arrays to obtain synchronized, multi-angle measurements, significantly enhancing detection accuracy with minimal added hardware. Machine learning models classify four key activities-“Falling,” “Walking,” “Lying,” and “Standing” -based on time series and statistical features, including the mean and variance from each antenna array. The system achieved a high classification accuracy of 98.3% using Ran-dom Forest models. Cross-environment testing further confirmed the model's robustness and generalizability, establishing this approach as a scalable, high-accuracy solution for real-time, location-independent activity recognition.