Real-Time Recognition of Human Postures for Human-Robot Interaction

To function in a complex and unpredictable physical and social environment, robots have to apply their intellectual resources to understand the scene in an efficient and intelligent way, similar to humans. Especially when interacting with humans, this cognitive task becomes more challenging. The work in this paper is focused on recognizing human actions and postures during daily life routines in real-time to understand human motives and emotions during a dialogue scenario. Using depth data, a real-time approach has been proposed that uses human skeleton joint angles to recognize 19 different human postures (standing and sitting). Feature vectors are constructed after pre-processing of joint angles. A supervised learning mechanism has been used to train the classifier using Support Vector Machine. Approximately 30000 training samples have been created for training purpose. The system recognizes all the postures accurately provided the skeleton tracker is working precisely when tested on the database. During live testing, the system reports 98.2% recognition rate, proving the potential of the proposed approach.