Intelligent Monitoring System for Elderly Living Alone Based on Multimodal Fusion and Cross-Modal Attention Mechanism

With the global aging crisis, existing health monitoring systems for solitary elders suffer from delayed response, high false alarms, and limited data diversity. This paper presents an indoor multimodal intelligent monitoring system integrating visual, acoustic, physiological, and environmental data via cameras, microphones, and wearables. The framework features improved ResNet-50 and temporal convolutional networks (TCN) for spatiotemporal behavior analysis, Wav2Vec 2.0 for speech modeling, and LSTM for sensor data processing. A multi-head attention-based dynamic weighting method explicitly models cross-modal dependencies, aided by lightweight adaptive-threshold feature selection for redundancy suppression. The anomaly detector combines LSTM-1D CNN to analyze temporal patterns with dual-validation (sensor cross-checking and temporal consistency) for false alarm mitigation. Experiments on our dataset show 2.1% fall detection false alarms, 2.3s emergency response, and >89% recall under low-light/noise conditions.