Head Gesture Recognition with Dense Net Ensemble Classifier from IMU Sensor Data

Abstract

Head gesture classification using Inertial Measurement Unit (IMU) sensor data has become increasingly important for touchless human–machine interaction, especially in wearable and assistive technologies. Recent studies show that nearly 65% of gesture-based interfaces still rely on vision-based systems, which are less reliable in low-light or occluded environments, while sensor-based interfaces improve accuracy by up to 20% in such conditions. This study introduces a novel framework for accurate head gesture classification using IMU data, focusing on robust preprocessing, effective data balancing, and advanced model design. Initially, the raw IMU signals are denoised and standardized through a dedicated preprocessing pipeline to ensure consistency across all gesture categories. To address the class imbalance commonly observed in real-world gesture datasets, the Synthetic Minority Oversampling Technique (SMOTE) is applied, which synthetically augments underrepresented gesture classes such as Jump Next Left and Stop Video, ensuring more balanced and representative training samples. The baseline models such as Perceptron and MLP are evaluated, revealing limitations in distinguishing subtle variations in gestures like MoveLeft_2s and MoveRight_2s, due to their shallow architecture and low expressive power. To overcome this, the proposed DenseNet-RFC framework is introduced, where DenseNet extracts deep spatial–temporal features from the IMU signal segments, and the Random Forest Classifier (RFC) robustly classifies these features across eight gesture classes: MoveRight_2s, MoveLeft_2s, Jump Next Right, Jump Next Left, Rise Volume-up, Rise Volume-down, Start Video, and Stop Video. This fusion of deep representation with ensemble classification leads to significantly improved accuracy, generalization, and interpretability of gesture recognition in noisy and imbalanced IMU data environments.