Factorized 3D-CNN for Real-Time Fall Detection and Action Recognition on Embedded System

We present a novel approach for skeleton-based action recognition and fall detection, optimized for real-time performance on embedded devices. Our method employs a factorized 3D convolutional neural network (3D-CNN) to efficiently extract spatiotemporal features from skeletal data. Initially, a 2D convolution layer is applied to capture spatial features from the input skeleton frames. Subsequently, a 1D convolution layer processes these spatial features to model temporal dynamics, effectively reducing the computational complexity compared to traditional 3D-CNN approaches. This factorization enables the creation of a lightweight model that maintains high accuracy while being suitable for deployment on resource-constrained embedded systems. Our approach is particularly advantageous for surveillance applications, such as autonomous driving or monitoring in elderly homes, where real-time action recognition and fall detection are critical for ensuring safety. Experimental results demonstrate that our model achieves high performance in recognizing various actions and detecting falls, highlighting its potential for practical real-world applications.