2024 年 37 巻 1 号 p. 12-21
Rotational motion estimation of a 3-D rotating object, i.e. estimating rotational posture and angular velocity at each time point from a sequence of images of the object is an important and challenging task. Previous research use feature extraction algorithms, but information extracted from these kinds of algorithms is not guaranteed to be optimal for rotation estimation. For this reason, as a method for taking the entire image as input and automatically extracting features from the given image, we use an image-based deep auto-encoder such that the latent variable can be interpreted as rotational representation by adding some constraints to the latent variable. Combining with Extend Kalman filter, we estimate not only rotational posture but also angular velocity and inertia ratio of the object. This method is validated using simulation data and it is shown that rotational motion can be estimated well. Also we explore ways to reduce the amount of labelled data used in the training dataset when training this model. Our findings indicate that the labelled data can be decreased to as low as 1% of the total training data without undermining the model's performance significantly.