Vision-Based Human Fall Detection Using 3D Neural Networks

The use of Machine Learning to monitor old people is crucial in providing immediate assistance and potentially life-saving interventions. With the rapid innovation in the field of Artificial Intelligence and Computer Vision, fall detection has seen significant improvements in accuracy and efficiency. Traditionally, 2D Convolutional Neural Networks (CNN) have been the main focus in fall detection research. However, these approaches have several drawbacks, including that 2D CNNs are primarily designed for spatial feature extraction and may not fully capture the temporal dynamics across multiple frames. This is because, for 2D CNN, the video frames are averaged out on time dimension within a time window. This project aims to explore and validate the use of 3D Convolutional Neural Networks (CNN) for fall detection, specifically in care home settings. The proposed 3D CNN keeps all frames in the time dimension (without averaging out video frames) and therefore can capture spatiotemporal dynamics of fall events more effectively, potentially enhancing detection accuracy. Experiment results indicate that the proposed 3D CNN achieved a G-Means, the geometric mean of recall and specificity, of 96.92%, an improvement of 1.9% over the 2D CNN.