Apparatus for recognizing activity in sports video using cross granularity accumulation module and method thereof

Abstract

The present invention relates to an activity recognition apparatus and method in sports videos using CGAM. According to the present invention, the apparatus for recognizing an activity in a sports video using CGAM (cross granularity accumulation module) comprises an object feature extraction unit that receives a video to be analyzed into a temporal attention module (TAM) in frame units to distinguish importance between frames, sequentially inputs the frames distinguished by importance into a plurality of convolution blocks and each spatial attention module (CBAM) placed between convolution blocks to output a first feature value, outputs a second feature value by a CGAM that generates an object representation by compressing different object information output from each spatial attention module, and extracts an object feature value by multiplying the first feature value and the second feature value; and an activity feature extraction unit that sequentially inputs the extracted object feature values into a recurrent neural network (RNN) and a fully-connected (FC) layer and classifies a final activity from a probability value for each activity estimated using a sigmoid function.

Dependencies are inferred from the claim text.