Activity Recognition using Kinect and Comparison of Supervised Learning Models for Activity Classification
Tanakon Sawanglok, Tananya Thampairoj, Pokpong Songmuang · 2018
This work presents a method to develop activity recognition using Kinect as a motion-sensing device and supervised learning for classification. Data from Kinect are continuous and independent frame representing in three dimensional axes from 20 human joints. The data then are trained for classify activities using supervised learning algorithms. The activities are 10 basic motion-gestures such as standing, waving, Thai-style greeting and walking. To compare supervised learning for classification in the task, four algorithms including neural networks, naive bayes, decision tree and support vector machine are applied to generate classification models. From experiment results, the best overall classification model was from neural network algorithm at about 75% accuracy while the second best was support vector machine with slightly lower accuracy. From analysis, the most incorrect activities were `wai' (Thai greeting) and `walking” in which were often misinterpreted to their similar activities as `bowing' and running', respectively.