ScholarWorks@성균관대학교

초록

Smart walkers that provide walking assistance for the elderly or disabled increasingly require the ability to recognize the condition of the user and realize real-time interactions. Among the core components of such systems, recognizing the walking intention and estimating the position of the user are essential. This study proposes a technique that simultaneously performs walking-intention recognition and position estimation using depth image sequences acquired from a single time-of-flight depth camera. For user-walking-intention recognition, the depth image sequence is used as the input to a transformer model. This enables the network to learn spatial depth features and temporal walking patterns in an end-to-end manner, allowing it to predict walking intention. For user-position estimation, pixel-level depth variation analysis based on depth differences is applied to quantitatively classify user movements. A Kalman filter is then employed to enhance the stability of distance estimation between humans and walking devices. The experimental results achieved an F1-score of approximately 74% for walking-intention prediction and an average position error of 0.723 ± 0.027 m for distance estimation. The proposed method can serve as a core perception and intention recognition technology for smart assistive-walker control. Future research will aim to further enhance the prediction performance of the developed system by integrating motion trajectory data from smart walkers.

키워드