Water Polo Covid-19 Analysis

May 20, 2021 | A study using computer vision to track the proximity of water polo players during a game

Data Collection

There is no published resource analyzing the x-y positional data of players in water polo and there is no current method of player localization and tracking in water polo. SportVU is a tracking technology developed in 2005 to track soccer matches in Israel that has expanded to use in professional basketball and other major soccer leagues. [1] It can collect the x-y positional data of players and referees during a game, in addition to a variety of other game statistics. NBA arenas with the system in use have six cameras, each responsible for a specific area of the court, capturing images at 25 times per second. [2]

SportVU has not yet expanded into the realm of water polo for two probable reasons. The first is the relative popularity of the sport compared to basketball in the United States or soccer in Europe. The software requires many high quality cameras which can be expensive to purchase and install. The other is the difficulties that arise when attempting to optically recognize athletes in a water polo pool, in which players frequently go under water and splashes block the view as players move through the water. Even the team of a player can be difficult to know if either their cap is shielded by the water or it falls off, as the cap is the only distinguisher between players of different teams. Players frequently go under water and splashes hide the positions of players and make it difficult to track the movements of players that are close together, even on high quality footage.

Due to the complex difficulties of water polo player tracking, the method of data collection for player localization used in this study was by user input by means of a computer vision guided user interface. Data was retrieved from a video posted to YouTube of a varsity high school water polo game at Kirkwood High School in St. Louis, Missouri between the two best high school teams in Missouri at the time, Parkway West High School and Kirkwood High School. [3] This non-professional video was optimal because, compared to other videos available, the camera is stationed at a fairly high vantage point, allowing for a view that is near-aerial. The camera is stationary and does not zoom, keeping 88% of the pool in the field of view at all times.

Games recorded professionally on high quality cameras feature zooming and panning which create a very clear image of the ball and the players near the ball, but fail to show the other players which are not near the play. For our analysis it was desired to have as much of the field of play in the view of the camera at all times. Because the video was produced non-professionally, it has poor video quality to the extent that no cap number on any player can be recognized even by a human. The identity of an individual player can only be known by a human if that player has a markedly different skin tone, body type, or swim stroke than other players on the team. More robust and developed systems like SportVU can identify individual players. The video used in this analysis, however, only allowed for the distinguishment of a player belonging to the home or the away team, and if that player was the goalies or one of the 6 field players on a team.

Before analyzing the game footage, the video was clipped to cut out timeouts and breaks between quarters once players had gone to their benches. The study was meant to analyze the proximity of players to each other while playing the game of water polo, not during breaks or when outside of the pool. Players on the bench were wearing masks, and teams did not switch sides at all during the game. Therefore if players were on the bench, they had no way of coming in close contact with a member of the other team.

In the first frame of the video, the user identified the positions of the players in the pool, distinguishing them as either an away field player, away goalie, home field player, or home goalie. After the first frame, the user skipped at most 100 frames before locating where these players had moved. The video had a frame rate of 30 frames per second, meaning that the locations of the 14 players were identified at least every 3.4 seconds for the entirety of the 45 minute 25 second video. For all of the unidentified frames, the positions of the players in the pool were interpolated based on the positions of those players in the user input frames on either side of the unidentified frame. At a rate of inputting at least every 3.4 seconds, with each frame requiring the input of the positions of 14 players, the entire video required over 11,000 data entries, although more were included at moments in the game which contained a large amount of dynamic movement.

If a player went off screen into the 12% of the pool that was not visible, the position was marked as unknown, and it was assumed that they were not within 6 feet of any other player. Again, due to the limitations of the one camera setup, only data for the positions of the players in the pool could be collected, leading to the assumption in the data that all 14 players in the pool never substitute, even though this is inaccurate. Due to the difficult method of data collection, only positional data for this game was collected.

Data Analysis

The x-y positional data of the players from the video was first manipulated by means of a homography, which transformed the pool as seen in the video into a rectangular 25 meter by 17 meter animated pool, the size of the pool field of play.

The x-y positional data of the players was then analyzed by means of the method established by [Knudsen et. al.] to analyze the spread of the Covid-19 virus during soccer matches in the Danish football league. After collecting the data, one player was chosen as the one infected with Covid-19. Over the course of the game, a danger zone existed as a circle of radius 1.5 meters extending from the center of the infected player. For every other healthy player, an exposure score was determined as the number of seconds that the healthy player was within the danger zone of the infected player. [Knudsen et. al.] treated the danger zone as having exponentially decaying effectiveness with a half-life of 2 seconds, meaning that if a healthy player stood for a second within what had been the danger zone 2 seconds prior, the healthy player would receive an exposure of 0.5 seconds rather than the typical 1 second had they been within the current danger zone. [Knudson et. al.] referred to this decaying danger zone as the tail following the infected player.

The summation of a healthy player’s exposure times gave the composite exposure over the course of a match for each player. After simulating the effect of one player having Covid-19, the simulation was repeated with a different infected player, and was again repeated until all players in the pool had been simulated as infected.

[Knudson et. al.] analyzed the positions of players individually, and could account for subbing and players leaving the field of play. Due to the limited availability of camera angles and player identification in our analysis, it was assumed that the same players on either team played the entire game. In water polo there is a goalie and six field players on each team, and although these six field players have specializations, they largely can end up playing any position in the pool, depending on their team’s need at a certain moment and the progression of the opposing team’s attack. Therefore, it made sense to only separate the exposure scores of the goalie, who stays at one end of the pool the entire game rather than swimming, from the exposure scores of the rest of the field players.

Results

The mean exposure score for the 45 minute 25 second piece of game footage (one game) for each field player was 147.6s, with a 95% confidence interval between 128.7s and 166.5s. The highest score for the field players was 504.8s, and the lowest score was 0s.

The mean exposure score for the goalies was 5.7s, with a 95% confidence interval between 1.7s and 9.6s. The highest score for the goalies was 47.9s, and the lowest score was 0s.