Real-time Recognition of Multi-category Human Motion Using μIMU Data

This paper describes a novel approach for human motion recognition via motion feature vectors collected from a micro Inertial Measurement Unit (μlMU), which measures angular rates and accelerations of the three different directions in the workspace based on MEMS sensors. The recognizer is composed of three parts. The first part is a preprocessor, in which Vector Quantization is used to reduce dimensions of vectors. Recognition is implemented by the second part, which is a classifier composed of Hidden Markov Model and an efficient second layer criterion. The third part uses a sliding window algorithm for precise recognition. There were 200 sequences (about 100,000 vectors) for 10 different kinds of motions tested in our work, including falling-down motion and other typical human motions. Experimental results show that for the given 10 different categories, correct recognition rates range from 95 %-100 %, of which the falling-down motion can be classified from others with a 100 % recognition rate.