Abstract
Human behavior analysis in indoor environments from ambient sensor data is one of the most challenging human-machine interaction research topics. In this paper, we present a deep learning-based novel approach for human posture recognition and emergency detection using an ultra-wideband radar sensor. The strength of this sensor is that it collects less privacy-related information than a regular camera. The sensor is installed on a mobile robot to observe a subject from a short distance. At first, raw data is collected from the sensor. Then, the data is used to train a Recurrent Neural Network (RNN) for modeling of different human activities and conditions, including normal and emergency conditions. Finally, the trained model is used for testing the model on new inputs. The experiments were performed using two datasets recorded in lab environments, and the proposed approach produced better recognition performance than the conventional ones. The proposed method can be applied in many prominent research fields (e.g., human-robot interaction) for different practical applications such as mobile robots for eldercare.