TLDR: A study investigated the accuracy of commercial computer vision and AI software for tracking player position, speed, and distance using broadcast footage from a 2022 FIFA World Cup match. Comparing data from three providers against a high-definition optical tracking system (TRACAB Gen 5), the research found that while the technology shows promise, accuracy is highly dependent on player detection, the type of camera feed (tactical feed being superior), and the provider’s specific AI models. Video resolution (720p vs 1080p) had no significant practical impact. Overall, speed tracking was more consistently accurate than position tracking, especially when players were undetected. The study recommends using tactical feeds to maximize player visibility and highlights the need for improved AI extrapolation models for undetected players.
In the fast-paced world of professional sports, understanding every move a player makes is crucial for training, strategy, and performance analysis. Traditionally, this has been achieved through advanced optical tracking systems or wearable devices. However, these methods often come with limitations, such as requiring specific stadium infrastructure or players wearing equipment, which can be restricted by sport regulations or personal preference.
A new frontier in player tracking is emerging with computer vision and artificial intelligence (AI) software that can analyze standard broadcast footage. This technology promises a more practical approach, as it requires no special equipment or setup by teams. But how accurate is it, and what factors influence its reliability?
A recent study titled “Concurrent validity of computer-vision artificial intelligence player tracking software using broadcast footage” by Zachary L. Cranga, Rich D. Johnstona,c,d, Katie L. Millse,f, Johsan Billinghame, Sam Robertsong, Michael H. Colea, Jonathon Weakleya,c,d, Adam Hewitta and, Grant M. Duthiebc, delved into these questions. The researchers aimed to determine if commercially available computer vision and AI software could accurately measure player position, speed, and distance using broadcast footage, and to understand how different camera feeds and video resolutions impact this accuracy. You can read the full paper here.
How the Study Was Conducted
The study collected data from a single match during the 2022 Qatar Fédération Internationale de Football Association (FIFA) World Cup. Three types of video feeds were used: the ‘tactical’ feed (a wide-angle view from a fixed camera at the halfway line), the ‘programme’ feed (what is televised to the public, including graphics and replays), and ‘camera 1’ feed (a slightly tighter view than tactical, without graphics). These feeds were analyzed at two common video resolutions: 720p and 1080p.
Three commercial tracking providers, all utilizing computer vision and AI, participated in the study. They analyzed the instantaneous position (x, y coordinates) and speed of each player. Their data was then rigorously compared against a highly accurate, high-definition multi-camera optical tracking system called TRACAB Gen 5, which served as the gold standard for accuracy.
Key Findings on Accuracy
The study revealed several important insights into the performance of AI-powered player tracking:
- Overall Promise, but Variability: Computer vision and AI software show potential for tracking players, but accuracy varies significantly depending on the provider’s specific processing techniques, the camera feed used, and to a lesser extent, video resolution.
- Player Detection is Paramount: A critical factor for accuracy is whether the software can ‘detect’ the player. When players were clearly detected, the accuracy for position and speed was significantly better. Errors increased substantially when players were ‘undetected’ (e.g., outside the camera’s view or obscured by other players).
- Camera Feed Matters: The ‘tactical’ feed, which offers a wide, elevated view of the entire pitch, generally led to better accuracy for both position and speed. This is because it maximizes player visibility and minimizes instances where players are out of frame or occluded. The ‘programme’ feed often resulted in poorer accuracy, especially when players were undetected.
- Resolution Not a Major Factor: While there was a statistically significant effect of video resolution (720p vs. 1080p) on validity, the practical differences in accuracy were not substantial. Both resolutions were deemed suitable, assuming appropriate AI models are implemented.
- Provider Differences: The study highlighted that the specific AI models and extrapolation techniques used by each provider play a crucial role. Some providers demonstrated superior ability to estimate player position and speed even when players were undetected, suggesting more advanced AI models.
- Position vs. Speed: While overall position accuracy (ranging from 1.68 to 16.39 meters RMSE) was not consistently suitable for detailed spatial tracking (like distance between players), it was much better when players were detected (0.44 to 1.14 meters RMSE for some providers). Speed accuracy (ranging from 0.34 to 2.38 m/s RMSE) was generally considered suitable across all configurations, even when players were undetected by some providers.
- Total Distance: The accuracy of total distance covered, a derived metric, also varied. Providers 1 and 2 showed suitable measures using the camera 1 and tactical feeds, while Provider 3 reported poorer validity. This was again linked to player detection rates.
Also Read:
- Sports AI Innovator Paradigma.ST Drives Business Growth with Advanced Computer Vision
- Unlocking Personalized Yoga: A Technological Vision for Better Health
Recommendations for Future Use
The researchers recommend that tracking providers primarily use a ‘tactical’ feed with either 720p or 1080p video resolution. This approach maximizes the number of frames where players are visible and detected, thereby improving overall accuracy. The study also emphasizes the need for continued research and development to refine the AI extrapolation models used by providers, especially for estimating player data when they are temporarily out of view.
While this study was limited to a single match and stadium, its findings offer valuable insights into the current capabilities and future potential of computer vision and AI in sports analytics. As AI technology continues to advance, the accuracy and reliability of player tracking from broadcast footage are expected to improve, making sophisticated performance data more accessible to teams worldwide.
