Terrain-Aware Head Gesture Recognition for Turret Control Using Helmet-Mounted IMUs and Vehicle Vibration Fusion

Head gesture-based control using inertial measurement units (IMUs) provides an intuitive alternative to conventional human–machine interfaces for mobile and vehicle-mounted systems. However, gesture recognition reliability degrades significantly under terrain-induced vibration and mechanically dynamic operating conditions. This study investigates terrain-aware head gesture recognition through the integration of helmet-mounted IMU measurements and vehicle vibration sensing to improve discrimination between intentional gestures and non-intentional motion artefacts. Vehicle vibration data were collected from a patrol vehicle traversing the Ndumo Border Patrol route, characterised by variable terrain roughness and dynamic excitation profiles. Triaxial seat-rack acceleration data were acquired at 10 kHz, anti-alias filtered and down sampled to 100 Hz before being integrated with IMU-derived head motion measurements using both vibration-aware data augmentation and early sensor fusion strategies. FFT-based spectral processing and classification using a lightweight fully connected neural network (FCNN) were implemented using the Edge Impulse framework. Experimental evaluation was performed using temporally independent training and testing segments to reduce overlap leakage and ensure realistic generalisation assessment. Results demonstrate that terrain-informed sensing substantially improves operational robustness under mobile conditions. The early-fusion approach achieved 98.45% independent-test accuracy under float32 inference and maintained 90.02% accuracy after int8 quantization, corresponding to an accuracy reduction of 8.43 percentage points. In comparison, the Clean IMU and vibration-augmented models exhibited reductions of 20.13 and 27.02 percentage points, respectively, demonstrating greater sensitivity to quantization. The evaluated models also exhibited low inference latency and compact memory requirements, supporting their suitability for real-time edge implementation. These findings demonstrate that treating terrain vibration as contextual information, rather than solely as environmental noise, can improve the quantization robustness and deployment characteristics of IMU-based gesture-recognition systems intended for mechanically dynamic platforms.