An efficient star skeleton extraction for human action recognition using hidden Markov models

This paper aims at finding an efficient approach for automatic human action recognition to classify human actions in both outdoor and indoor environments. A human action recognition system (HARS) collects video frames of human activities, extracts the desired features of each human skeleton. These characteristics are calculated, classified to build a skeleton database that can distinguish almost human gestures. This HARS converts every sequence of human gestures to the sequences of skeletal joint mapping (SJM). Then it assigns corresponding observation symbols to each SJM. Those observation sequences are used to train of hidden Markov models (HMMs) corresponding to seven actions: standing, walking, running, jumping, falling, lying, and sitting. Baum-Welch and forward-backward algorithms are employed to find optimal parameters of each HMM. During a recognition phase, each human gesture sequence is converted to an observation sequence and put into seven optimized HMM models. The current action can be identified by finding a model with the highest probability. The experimental results show that the proposed HARS offers high accuracy of action recognition in real-time.