AUTOMATIC ANALYSIS OF TIMING AND SYNCHRONY IN ROWING

Abstract

Synchrony and Timing are key components of the sport of rowing. This thesis investigated different ways of determining the synchrony between two athletes on rowing ergometers in a video. First, human pose estimation techniques were used to extract body key points and compare the synchrony of one or multiple body parts across the length of the video. Second, a video classification model based on either the R3D or I3D model were trained to classify rowing videos as synchronized or unsynchronized and their performance was compared. The pose estimation approach could detect and highlight small asynchronies between rowers that are hard to detect for humans and could provide valuable additions to rowing training. For the video classification approaches, the R3D-based model reliably classified rowing videos and generalized well to unseen and slightly different data. It outperformed the I3D-based approach, which performed well on the training data but generalized poorly to different data.