昆虫視覚系の運動検出器モデルとEcho State Networkを用いた人の動き分類

抄録

Spatio-temporal information is an important cue for action classification, whereas the computational cost for handling spatio-temporal information in neural networks is very high. In this study, we developed a action classification algorithm that processes spatio-temporal information efficiently by combining the elementary motion detector (EMD), which is a low-computational-cost neuro-inspired motion sensing algorithm, and the echo state network (ESN), which is a three-layered neural network with a recurrent layer. We evaluated the algorithm by classifying six kinds of human action, and compared the accuracy of the proposed algorithm with that without preprocessing of the EMD to clarify the effect of preprocessing. The results showed that the algorithm was successful in classifying the human action and that the accuracy was increased by the EMD drastically.