Human action recognition using depth motion map and Resnet ☆

Human action recognition is an active research topic in recent years due to its wide application in reality. This paper presents a new method for human action recognition from depth maps which are nowadays highly available thanks to the popularity of depth sensors. The proposed method composes of three components: video representation; feature extraction and action classification. In video representation, we adopt a technique of motion depth map (DMM) which is simple and efficient and more importantly it could capture long-term movement of the action. We then deploy a deep learning based technique, Resnet in particular, for extracting features and doing action classification. We have conducted extensively experiments on a benchmark dataset of 20 activities (CMDFall) and compared with some state of the art techniques. The experimental results show competitive performance of the proposed method. The proposed method could achieve about 98.8% of accuracy for fall and non-fall detection. This is a promising result for application of monitoring elderly people