Divers have developed a specific gesture language for efficient communication underwater. As companion robot drones are increasingly used to help divers, it is essential that these robots can understand basic commands. This paper focuses on diver gestures classification using geometrical features extracted from the diver's upper limb movements on underwater RGB-video. The extraction of upper limb skeleton points relies on pre-existing skeleton detection algorithms. Three geometric features are examined: angles, joint-line projections, and joint-joint distances. The classification is conducted using a 3-layer Bi-LSTM neural network for 11 different gestures. The database is acquired in a pool with 4 subjects and contains a total of 200 hand-cut videos. The recognition results show an accuracy of about 50% for the anangles and joint-lines features, this validates the potential of the method based on 2D skeleton extraction in videos.This study shows that the classification results are strongly related to the type of geometric feature used and to the characteristics of the gesture.