A child caring robot for the dangerous behavior detection based on the object recognition and human action recognition

Abstract

In this paper, a child caring robot is developed for detecting some dangerous behavior performed by child in the domestic environment based on the human action recognition and object recognition technologies. A human behavior is an interactive process between human and objects. Therefore, three factors need to be considered: the engaged objects, human actions and the relationship between human and the engaged objects. In our application scenario, a correlative filter is proposed to improve the stability of the object recognition with human interference. For the human action recognition, a new motion encoding method by using the Euclidean distant matrix (EDM) between joints is introduced and a convolutional neural network is utilized. Evaluation on the Northwester-UCLA dataset verified the effectiveness of this method when action categories are small. The proposed action recognition method is simple and efficient, which is crucial for online behavior detection. Extensive experiments in the real physical world for detecting the behavior of eating allergic fruit and touching/playing with electrical socket have achieved good performance.