Classification of Human Posture with RGBD Camera: Is Deep Learning Necessary?

Abstract

AbstractWe describe an approach to human posture classification using RGBD camera (Kinect V2 sensor) data. We compared deep learning methods for human posture classification versus classical data classification methods. We conducted a user study where participants assumed various postures, including whole body, upper and lower body, as well as body transition motion. Several classical data classification methods, such as support vector machine, random forest, neural network, and Adaboost, were used for posture classification. Results show that the posture classification accuracy for the classical data classification methods is between 75% an 99%. The accuracy of the classical data classification methods is comparable to the accuracy of the long-short-term-memory (LSTM) deep learning method which is between 86% and 99%. Our findings suggest that the use of the classical data classification methods on the RGBD camera data is likely sufficient for posture classification, at least for certain task scenarios without incurring the overhead of deep learning.KeywordsComputer visionHuman postureDeep learningMachine learning