A hybrid HMM/SVM classifier for motion recognition using μIMU data

This paper describes a novel approach for human motion recognition via motion features extracted from sensor data. The classification process consists of two phases. The first one is a preprocessing of raw signals. Median Filter is used to filter pulse noise while vector quantization is used for Gaussian noise and reducing dimensions in this phase. The second one consists of a hybrid HMM/SVM classifier. Outputs from the first phase will be estimated by different pre-trained HMMs, and the results of the likelihood will be classified by the SVM classifier to identify the motion. With data collected from the mulMU equipment, falling-down motion can be told from non-falling- down motions with a correct recognition rate better than 99%. When the SVM training samples are labeled carefully and chosen bias, 100% correct recognition rate can be reached. The algorithm proves robustness and accuracy.