Blogs

Computational models applied to football calculate player orientation and predict the most feasible pass

(Original UPF news)

Data science in the world of football has advanced in leaps and bounds in recent years, and the top squads already have technical analysis teams working with new metrics to try to optimize the performance of the team. “Broadly speaking, these statistics come from player tracking data, which indicate the position of the players on the field at every moment in time. Based on tracking, predictive models can be built or similar moves can be grouped on the basis of patterns produced by the interaction of the players on the field”, states Adrià Arbués, UPF PhD student with the Image Processing Group (GPI) at the UPF Department of Information and Communication Technologies (DTIC).

But these predictive models, some of which are already used today, still lack some specific variables that can dramatically affect the outcome of the play, such as player orientation. In the current version of football, orientation is a vital skill that players need to master in order to make/receive a decent pass and stay ahead of the defence. However, tracking data do not include this type of information as they do not yet include pose models. 

“Thus, our recent research has involved creating two models: the first finds player orientation according to their body position and the second predicts which pass has the greatest/least risk at any given time, thanks to the orientation of all of the players”,  Arbués explains.

This research has been carried out by researchers at Pompeu Fabra University and FC BarcelonaAdrià Arbués-Sangüesa, a PhD student with the GPIAdrián Martín, a researcher with the Image Processing for Enhanced Cinematography Research Group (IP4EC), Coloma Ballester and Gloria Haro, coordinator and member of the GPI, respectively, all of them at the (DTIC), with Carlos Rodríguez and Javier Fernández (FCB).  

Calculation of the orientation of the player's body based on body position

The first part of the research was presented in an article that was recently accepted by the International Conference on Image Processing (ICIP2020). In this paper, the authors calculate player orientation based on their body position; more specifically, using a position model called OpenPose and, mapping the parts corresponding to the upper torso (hip and shoulders) in a two-dimensional plane, “we can draw a normal vector on the plane, which will give us the above orientation, in cases in which we cannot identify the hip or shoulders, the position of the facial parts plays an important role”, Arbués adds.

Achieving a refined estimation of player orientation

“The image quality may not be as desired in certain cases, and players who are away from the camera have a significantly lower resolution; hence, we try to refine the estimation of orientation in two ways: using a model of machine learning that tells us whether the player is facing or back to the cameraoffsetting orientation with the position of the ball, since most players tend to be slightly biased by its position”, Arbués, first author of the work, continues to explain.

“Thanks to the information supplied by Futbol Club Barcelona, we managed to compare the orientation obtained with our method with real match data, captured using wearable sensors; our method has a median error of less than 30 degrees”, Arbués points out. 

Which pass is most feasible at any time?

In the second part of the research, the researchers define a new metric that measures the feasibility of pass events; that is, at a given time, when a pass is about to be made, which player is best suited to receive the ball?

“To answer this question, we have combined three independent pass feasibility estimations, based on: (a) the distance between the passer and potential receivers, (b) the defensive pressure on the passer/receiver, and (c) the compatibility of orientations between passer and receiver, obtained by means of a geometric solution”, Arbués states. 

Thus, “we show that the latter is the key piece or ingredient to get the most likely pass at those moments of the match. To estimate it, we chose a geometric solution in which cones are defined with each player’s the field of vision (defined by orientation) and potential intersections are sought between the cone of the passer and that of potential receivers”, first author of the works adds.

Using a database of over 6,000 passes, the results show that the model is sufficiently robust, since 70% of the passes made are among the top 3 most feasible in our model and that orientation is crucial for decision-making in most cases (player position, area of the field).

Related works: