Comparing Recognition Methods to Identify Different Types of Grasps for Hand Rehabilitation
Beatriz León, Angelo Basteris, Farshid Amirabdollahian · University of Hertfordshire Research Archive (University of Hertfordshire) · 2014
Abstract—Grasping activities are extremely frequent in the set of activities of daily living. This causes severe impairments for stroke survivors, whose wrist and hand may suffer from a variety of symptoms such as spasticity, hypertone and muscular weakness. Intensive repeated movement performance is at the base of robot-therapy. Thus, patients may benefit, in terms of functional recovery, from the integration of grasp gestures in robot mediated exergaming. In this feasibility study, we developed and tested three methods for recognizing four different grasp postures performed while wearing an exoskeleton for hand and wrist rehabilitation after stroke. The three methods were based on the statistics of the produced postures, on neural networks and on support vector machines. The experiment was conducted with healthy subjects, with no previous injuries on the hand, during grasping of actual objects and then repeated using imaginary objects. We compared the three methods in terms of accuracy, robustness with respect to the size of the training sample, inter-subjects ’ variability, differences between different postures and evaluating the presence of real objects. Our results show that the support vector machine method is preferable in terms of both accuracy and robustness, even with a small training sample, with training times on the order of milliseconds. Keywords—grasp posture recognition; stroke rehabilitation; Sup-port Vector Machines; Neural Networks. I.