A Household Tool Classification Method Based on Parts Affordance Semantic Combination

To meet the needs of machine intelligence for tool affordance cognition in human-robot coexisting-cooperative-cognitive environment, a clustering method based on tool part affordance semantic combination, inspired by the human bottom-up cognition patterns, is designed for household tool representation and modeling. Firstly, an edge representation method of tool affordance parts is designed and modeled based on structured random forest. Then, a whole tool representation method combined with saliency degree of each tool part in high-level semantic space is designed based on the idea of affordance parts combination, and the tool affordance dictionary is constructed through clustering. In the online detection phase, the saliency degrees of affordance parts of the test samples are combined, and the classification of household tools is determined according to the distance residual between the test sample and the tool affordance dictionary. In the experiments, 7 kinds of affordance parts are combined and clustered into 5 kinds of tools. When different kernel values are selected for various tools, the classification accuracy is more than 90%. Even if the kernel value is fixed as 3, the classification accuracy is more than 85%. The experimental results show that comparing with traditional feature-based representations, the addition of affordance semantic deepens the robot understanding for tool affordance, the accuracy and efficiency of household tool classification based on this combined affordance part dictionary are improved significantly, meanwhile the affordance similarity estimation between the tools and the search of the optimal substitute tool are also realized.