A wearable and non-wearable approach for gesture recognition

A natural way of communication between humans are gestures. Through this type of non-verbal communication, the human interaction may change since it is possible to send a particular message or capture the attention of the other peer. In the human-computer interaction the capture of such gestures has been a topic of interest where the goal is to classify human gestures in different scenarios. Applying machine learning techniques, one may be able to track and recognize human gestures and use the gathered information to assess the medical condition of a person regarding, for example, motor impairments. According to the type of movement and to the target population one may use different wearable or non-wearable sensors. In this work, we are using a hybrid approach for automatically detecting the ball throwing movement by applying a Microsoft Kinect (non-wearable) and the Pandlet (set of wearable sensors such as accelerometer, gyroscope, among others). After creating a dataset of 10 participants, a SVM model with a DTW kernel is trained and used as a classification tool. The system performance was quantified in terms of confusion matrix, accuracy, sensitivity and specificity, Area Under the Curve, and Mathews Correlation Coefficient metrics. The obtained results point out that the present system is able to recognize the selected throwing gestures and that the overall performance of the Kinect is better compared to the Pandlet.