A weighted-fusion assisted monitoring system for disabled people leveraging EfficientNet and VGG16 collaboration

The intensification of global aging and the continuous expansion of the disabled population have put forward an urgent need for their daily safety supervision. Traditional single-mode monitoring systems suffer from poor real-time performance and high false alarm rates, making it difficult to cope with multiple security threats in complex scenarios. To this end, a disabled safety monitoring model based on an improved Efficient Neural Network (EfficientNet) and a Visual Geometry Group 16 (VGG16) layer network is proposed. This model has improved the performance of the underlying network through targeted improvements: SiLU replaces the Swish activation function in EfficientNet and is embedded with non-local modules, enhancing the modeling ability of remote spatiotemporal dependencies and achieving accurate real-time recognition of abnormal behaviors such as falls; The introduction of SENet channel attention mechanism and SiLU replacement of ReLU activation function in VGG16 enhances the spatial feature extraction and recognition ability of VGG16 for environmental risks such as fires and obstacles. The core innovation of the research lies in the design of a weighted feature fusion strategy, which effectively achieves cross-modal alignment and collaborative analysis of behavioral temporal features and environmental spatial features. Validation using the UR Fall Detection dataset showed that compared to the unimodal system, the false positive rate and false negative rate of the bimodal system were reduced by 42.98% and 60.42%, respectively. The accuracy rate and F1 score reached 96.37% and 95.92%, respectively, with a detection latency of 37.81 ms. Additionally, the recognition accuracy for sudden abnormal behaviors and environmental risks both exceeded 90%, with time latency and memory consumption reduced by 37.89% and 30.26%, respectively. In addition, experimental results on a custom dataset (Real Disability Safety Dataset, RDSD) showed that the accuracy, recall, and F1 score of the system reached 92.15%, 89.42%, and 90.76%, respectively. These performance data were significantly better than security monitoring models based on Vision Transformer and Time Convolutional Networks, as well as security monitoring models based on Deep Convolutional Long Short-Term Memory Networks. The proposed system has distinct advantages in detection accuracy, robustness, real-time performance, and edge adaptability through heterogeneous model collaborative optimization and edge lightweight design, providing a high-precision and robust solution for disability monitoring.