USING TEMPORAL CONVOLUTIONAL NETWORKS FOR POSSESSION ESTIMATION IN FOOTBALL

The use of tracking data in the field of sport analytics has increased in the last years as a starting point for detailed analyses about several aspects of a match. In parallel, Temporal Convolutional Networks have established themselves as a powerful way to process sequential data: TCNs have outperformed RNNs on different tasks and are based on spatiotemporal convolutions, enabling to exploit the improvements in parallelization technologies. This work uses TCNs to classify the status of a football match, which can take on one of three possible values: inactive game, ball owned by the home team, ball owned by the away team. The dataset consists of tracking data collected during the 2019-20 season of the Italian Serie A. In analogy with the state-of-the-art, the data are processed in the TCN using 1×k filters, so that for each object an independent trajectory embedding is built, i.e. a fixed-size representation of the movement of the object over time. However, since there is no explicit ordering among the objects in the tracking data, it is necessary that the network has a permutation-invariant layer Λ (Lambda), so that the prediction about the game status doesn’t change if e.g. two players from the same team are swapped. The presence of the layer Lambda allows then to transform the set of trajectory embeddings into a tensor that can be processed by a Feed-Forward Neural Network to produce the final classification. Three different alternatives are proposed for achieving permutation invariance: an element-wise sum over the embeddings, a self-attention module and a TCN that uses 2D convolution instead of 1×k filters. On the other hand, the possibility is explored of splitting the classification into two separate predictions, telling respectively if the game is active and – if so – which team owns the ball. This can be done either by means of two network branches sharing the same TCN or by means of two separate networks, each with its own TCN. From these considerations it is possible to derive 17 candidate architectures, differing in the number of branches/networks and in the Lambda layers within them. In the evaluation phase the performance of these models is compared in terms of overall accuracy. Furthermore several issues are discussed concerning the impact of data quality on the accuracy, the use of ablation studies to investigate which parts of the input data concur to the prediction, and the types of errors made by the network in the testing phase.