Design and Evaluation of ADANet: A High-Fidelity Motion Acquisition Framework for Assistive Gesture-Based Interfaces

Accurate and low-latency gesture recognition is critical for real-time assistive technologies that enable individuals with motor impairments to interact more intuitively with their environment. However, current systems often suffer from poor signal fidelity, limited adaptability, and high computational overhead. To address these limitations, this article presents ADANet — Advanced Disability Assistive Neural Network-driven framework for accelerometer-based gesture recognition, tailored for wearable assistive applications. The proposed system combines a high-fidelity ADXL335 triaxial accelerometer with an ESP32 microcontroller, forming a lightweight and cost-efficient motion acquisition pipeline. A structured preprocessing architecture is developed, incorporating zero-lag Butterworth filtering, entropy-based temporal smoothing, and computation of 33 handcrafted statistical features, including axis-specific jerk, signal magnitude area (SMA), and range-normalized entropy. ADANet’s compact yet expressive neural architecture is optimized through comprehensive ablation studies to ensure strong generalization with minimal latency. Our model was trained on data from 21 healthy participants(12 male & 9 female) performing eight functional gestures demonstrate a test accuracy of 84.47%, F1-score of 0.84, and AUC $\geq 0.96$ , outperforming traditional classifiers such as XGBoost, SVM, and KNN across multiple evaluation metrics. This work validates a robust end-to-end instrumentation-to-inference framework and establishes ADANet as a practical solution for embedded gesture recognition in assistive devices, mostly for disabilities, where signal integrity, adaptability, and inference efficiency are jointly critical.