Classifying bio-inspired model of point-light human motion using Echo State networks

Abstract

We introduce a feature extraction scheme from a biologically inspired model using receptive fields (RFs) to point-light human motion patterns to form an action descriptor. The Echo State Network (ESN) which also has a biological plausibility is chosen for classification. We demonstrate the efficiency and robustness of applying the proposed feature extraction technique with ESN by constraining the test data based on arbitrary untrained viewpoints, in combination with unseen subjects under the following conditions: (i) lower sub-sampling frame rates to simulate data sequence loss, (ii) remove key points to simulate occlusion, and (iii) include untrained movements such as drunkard's walk. © Springer International Publishing AG 2017.