Action Recognition of Excavators Using Physical Simulator and Real Image Data with Class-Dependent Data Augmentation.

In this study, we proposed a training method using joint points obtained from physical simulations and real image data for the action recognition of excavators. The proposed method classifies the action recognition of excavators into position detection, skeleton detection, and action recognition models. The first two models are trained using the real image data, whereas the action recognition model is trained using the joint point data obtained from the physical simulation. For the action recognition model, we proposed a data augmentation method based on the features of the actions of the excavator. Experimental results indicate that the proposed method can achieve better accuracy than the conventional method that uses real video data, though the proposed method does not use any real video data for training.