Fall detection algorithm based on lightweight 3D residual network

Aiming at the many parameters and high computational complexity of video-based deep learning fall detection models, we propose a lightweight fall detection algorithm for 3D residual networks. In this approach, we design a low-rank depth-separable convolution structure. When performing deep convolution, the 3-dimensional parameter matrix is decomposed into 1-dimensional and 2-dimensional parameter matrices to reduce the model parameters and thus improve the performance. Meanwhile, the dataset is built by referring to the format of the URFall dataset and capturing videos of human falling and non-falling states from multiple angles using RGB cameras. The experimental results show that the lightweight 3D residual network can achieve 98.23% accuracy in distinguishing falls from non-falls, and the sensitivity and specificity are kept at a high and stable level.