The automatic analysis of different sports spreads increasingly, because such systems can help the athletes, especially in competitive sports. In the past decades vast amount of visual information is gathered about sport events, trainings and tournaments. However the utilization of this information hasn’t started yet in an organized, consequent way except some professional clubs of popular ball games. Analyzing the image content may help improve the motion patterns, thus could have a great impact in many sport disciplines, such as in rowing.
The goal of this work is to design a system, which can identify the major body parts of the human body (head, hip, knees, ankles, elbows and hands) in video sequences taken from side view of athletes, practicing on indoor rowing machines. Further target is presenting and visualizing the captured data in a well understood format and deriving useful numerical parameters (e.g. stroke rate, length of segments), in order to support the coaches’ decisions.
In this work I review the existing methods in the field, present my former method and demonstrate the motivation behind using a new technique, the neural network based learning. I have collected and annotated video data for training, then designed, implemented and trained a convolutional deep neural network. I also present detailed results of the method and comparison against existing general human pose estimation solutions.