Exploratory: Sports Data Science

A new work developed within the Sports Data Science exploratory of SobigData++ has been presented at the 19th edition of the European Conference of Machine Learning and Principles and Practice of Knowledge Discovery in Databases (ECML-PKDD), one of the prominent European conferences dedicated to data mining and machine learning.

In recent years, several companies have been involved in sports analytics through the collection of spatio-temporal data describing the events that take place during soccer matches. These data are collected manually by human operators, constituting a considerable cost in terms of time and economic resources. Most of the work in the literature focuses on methods of video summarization, which can detect the salient but infrequent actions that occur in soccer matches, such as goals, replays, highlights, and play-breaks. Nevertheless, we are still far from the construction of a model that can automatically recognise the most important or frequent event in soccer, such as passes, shots, fouls, offsides, etc.

PassNet is a method to recognize the most frequent events in soccer, i.e., passes, from video streams. Such a model combines a set of artificial neural networks that perform features extraction from video streams, object detection to identify the positions of the ball and the players, and classification of frame sequences as passes or not passes.

PassNet processes input video (a) combining two convolutional neural network (c and b) to feed a recurrent neural network LSTM module (e).

The architecture of PassNet, shown in the figure above,^[1] combines three tasks: 1) feature extraction reduces the dimensionality of the input using ResNet18; 2) object detection detects the players and the ball in the video frames using YOLOv3; 3) sequence classification classifies sequences of frames as containing a pass or not using a Bi-LSTM.

Reported results highlight how PassNet significantly outperforms several baselines on four challenging scenarios designed to test the pass detection capability, indicating that the approach is able to learn from data to annotate pass events.

Screenshot from a video showing how PassNet annotates match US Sassuolo vs. FC Internazionale (Italian first division) as the match goes by. The left side shows the match, a label “Pass” appears every time a pass is detected. The right side shows two animated plots comparing the real (red) and the predicted (blue) labels.