Dual-Module Vision-Based Framework for Close-Proximity Real-Time Fall Detection

Falls pose a serious threat to older adults' independence and well-being, with sit-to-stand (STS) transition frequently associated with fall incidents. To address this challenge, we propose a real-time fall detection integrated into a mobile assistive robot SkyWalker. Our approach utilizes a depth camera with onboard processing capabilities positioned at close proximity (approximately 0.5 m). A 3D skeletal model derived from MediaPipe tracks the user's motion in real-time, extracting 14 key kinematic features that capture biomechanical information. These features serve as input to a dual-modular classification framework based on support vector machines (SVMs): one classifier predicts STS phases (sitting, rising, switching, standing). while the other identifies irregular motions indicative of falls. By focusing on this reduced yet discriminative feature set, our system remains both computationally efficient and robust to skeleton distortions often encountered at close range. We evaluated our approach using separate datasets for phase classification and fall detection, achieving high accuracy in real-time classification. Future work will extend the system to enable proactive fall prevention strategies, ensuring safer STS transitions for older adults.