AI-baserat kamerasystem för detektionav fall hos äldre i hemmiljö

Sammanfattning

This thesis presents a camera-based artificial intelligence system for fall detection in home environments, with elderly users as the intended application area. The aim is to enable real-time fall detection without storing video data, addressing privacy concerns. The prototype uses YOLOv8-pose to extract body keypoints from video frames and a Long Short-Term Memory network to predict fall versus non-fall sequences. A custom dataset of 114 videos was recorded in two home-like environments, covering six fall scenarios and seven activities of daily living. The model was trained with video-level data splitting and a multi-seed evaluation using five seeds. A parallel sequence-level split was tested to study the effect of data splitting on performance. With video-level splitting, the model achieved an average F1-score of 0.680, recall of 0.846, precision of 0.569 and an area under the receiver operating characteristic curve of 0.910. Sequence-level splitting produced substantially higher results, with an F1-score of 0.985, indicating that this approach can overestimate performance due to information leakage. In a real-time test using a webcam and a threshold of 0.7, all five staged falls were detected, while one of eleven activities of daily living caused a false alarm. The measured latency from first fall prediction to alarm was 0.42 ±0.07 s. The results show that pose-based fall detection is feasible as a real-time prototype, but a larger and more diverse dataset is needed to improve precision, robustness and generalization before deployment.