A novel dual-arm motion discrimination method using recurrent probability neural networks for automatic gesture recognition
Yuki Hiramatsu, Taro Shibanoki, Keisuke Shima, Toshio Tsuji · 2011
As gestures are mainly characterized via combinations of left and right arm movements, automatic gesture recognition requires accurate identification of separate arm motions. This paper proposes a novel dual-arm motion discrimination method that combines posterior probabilities estimated independently for left and right arm movements. In this approach, the posterior probabilities of individual single-arm motions are first estimated from measured biological signals using recurrent probabilistic neural networks. Then, the estimated posterior probabilities are combined automatically based on the motion dependency that exists between the arms, making it possible to calculate the joint posterior probability of dual-arm motions. With this method, all dual-arm motions consisting of individual single-arm movements can be discriminated through learning of single-arm motions only. In the experiments performed, the proposed method was applied to the discrimination of 15 dual-arm motions made up of three movements for each arm. The results showed that the method enables high discrimination performance based on learning of only three motions for each arm (average discrimination rate: 98.80 ± 0.68%).