Frames Extraction from Table Tennis Competition Videos for Action Classification Using Optical Flow and Fuzzy Rules

Abstract

To recognize actions using a neural network model, it is necessary to extract the correct frames from the video for the input of model. Extraction of frames is an important issue that could be poor recognition results or costs computation time. This study proposes a new extraction method that combines optical flow and fuzzy rules. First, optical flow is used to calculate the values of the x and y vectors of the motion in consecutive frames. After expert discussion, rules are formulated to define the optical flow values for each action as fuzzy semantic words and stored as a fuzzy rule base. For the experiment, serving action is further subdivided into tossing, hitting and receiving parts in table tennis video. Using fuzzy rules based on the x and y optical flow values of different actions, the current action type can be determined, and action frames can be extracted more accurately, improving the accuracy of table tennis action recognition, the final result of table tennis action recognition reached up to 69.8% accuracy.