AI-based human activity classification using BHI360

Falls are a leading cause of injury-related hospitalisations among adults aged 65 and over, often resulting in serious injuries and loss of independence. With rapidly aging populations, there is an urgent need for continuous, unobtrusive fall detection systems that can operate autonomously on wearable devices. This thesis presents an end-to-end pipeline for embedded fall detection using the Bosch BHI360 smart sensor—a wearable platform with a 6-axis IMU and 256 kB RAM. The system addresses the deployment gap between high-performing research models and practical, always-on wearable implementations by designing for embedded constraints from the outset. Thirty recording sessions were conducted with a healthy volunteer wearing the BHI360 sensor on the left wrist. Each session captured diverse activities including motion, postural transitions, static periods, and simulated falls. Tri-axial accelerometer and gyroscope data were sampled at 200 Hz and processed using 2-second windows with 0.25-second strides, from which handcrafted features were extracted. Two classification models were evaluated: a HistGradientBoostingClassifier (HGB) and a DecisionTreeClassifier (DT). Using leave-one-session-out cross-validation, the HGB model achieved the highest performance (macro recall = 0.84, accuracy = 0.82). The DT model achieved competitive performance (macro recall = 0.81) and was successfully converted to C code using m2cgen, integrated into the Bosch SDK, and deployed as firmware on the BHI360 device. The final system demonstrates that machine learning-based fall detection can be effectively deployed on embedded platforms when guided by hardware-aware design principles. This work establishes a reproducible framework for embedded AI research and contributes toward practical, scalable health monitoring systems.