Visual aided training of billiards based on depth estimation

Billiards is a popular recreational sport, but because of the strong technical movement, beginners can not choose the relevant shot path. Given this situation, it is very promising and meaningful to use computer vision to assist billiards. However, due to the limited height and size of billiard ball, mobile phone photography can only get a side view, unable to accurately convey the cue ball's relative position and other balls. Therefore, the white ball hitting path planning can not achieve the ideal hitting effect in reality to assist users. This paper proposes a method of calculating the relative position of the cue ball and other balls by using monocular depth estimation and restoring them to the top view. Traditional monocular depth estimation requires the artificial assumption that the relationship between RBG image and depth satisfies a certain parameter model, which results in limited prediction accuracy and consumes computational resources. In this paper, the monocular prediction method based on deep learning is adopted, and the combination of PVCNN and global feature extraction method is proposed, which can improve the prediction accuracy and reduce memory consumption. The experimental results show that this method can be used to estimate the depth of information of billiards, which lays a foundation for the visual-assisted training of billiards based on mobile phones.