Fall detection based on CNN models implemented on a mobile robot

Fall accidents are serious events that need to be addressed. Generally, elderly people could suffer these accidents that may lead injures or even death. The use of Convolutional Neural Networks (CNN) has achieved the state of the art for fall detection, but it requires a high computational cost. In this work, we propose an efficient CNN architecture with a reduced number of parameters, which is applied to fall detection in a service with a mobile robot, equipped with a resource-constrained hardware (Nvidia Jetson TX2 platform). Also, different pre-trained CNN models are compared to measure their performances in real scenarios, in addition with other functions like following people and navigation. Furthermore, fall detection is carried out by extraction of temporal features obtained with an Optical Flow extraction from two consecutive RGB images. The proposed network is confirmed by our results to be faster and more suitable for running on resource-constrained Hardware. Our model achieves 88.55% of accuracy using the proposed architecture and it works at 23.16 FPS on GPU and 10.23 FPS on CPU.