Quaternion-Based Human Gesture Recognition System Using Multiple Body-Worn Intertial Sensors

In this study, we designed a multi-sensor gesture recognition system that can classify among six different human gestures. Data was collected from eleven participants using five gyroscopic motion sensors tied to their upper body. A total of 1080 samples were collected, which contain almost 6000 gestures collected within a span of 90 minutes. The data were processed and fed into a multiclass Pattern Classification system to classify the gestures. We trained Support Vector Machines and Artificial Neural Networks on the same dataset under two different scenarios to compare the results. A similar study was performed before using modified Hidden Markov Model but the data was collected using a single sensor. Our study indicates that near perfect classification accuracies are achievable. However, such accuracies are more difficult to obtain when a participant does not participate in training even if the test set does not contain any data from the training set.