From Controlled Data to Real-World Conditions: Vision-Based Fall Detection for Elderly Monitoring

Falls are a major threat to independent living among older adults, yet reliable automatic fall detection in real homes remains difficult because of the technical challenges such environment represent, and privacy concerns associated with constant monitoring. This project develops and evaluates a pose-based fall detection system for AI Care in which human motion is represented by sequences of body landmarks. Seven public datasets were pooled and harmonized into a common skeleton representation using MediaPipe, while two additional datasets were reserved entirely for out of distribution evaluation. Five temporal and spatio-temporal architectures were compared: BiLSTM, Transformer, temporal convolutional network, spatial-temporal graph convolutional network, and CNN+LSTM. Ablation studies on feature dimensionality, temporal pooling, and sequence length, and perturbation experiments assessed robustness to realistic input degradation. A real-time pipeline combining person detection, identity tracking and temporal buffering was also implemented. CNN+LSTM using 3D position achieved a mean accuracy of 0.942 and an F1-score of 0.940 on the test set. On the two unseen datasets, the model obtained F1-scores of 0.928 and 0.654, revealing both transfer to new conditions and an important generalization gap. The principal errors originated from the absence of a floorrelative reference, under-represented viewpoints and activities, severe occlusion, and subjects leaving the camera view. Live multiperson tests further showed that motion blur can cause detector and tracking failures that reset the temporal sequence before classification. Overall, the findings establish skeleton-based CNN+LSTM classification as a promising approach to fall recognition, while emphasizing that strong results on segmented clips do not guarantee reliable real-world operation. Practical deployment will require scene-relative spatial information, training data that better represent a realistic household