A Computer Vision Approach for Pass Detection on Soccer Broadcast Video

The annotation of the events that occur during a soccer match is a primary issue for companies that produce data for analytical purposes. Nowadays, the annotation is mostly manual, i.e., humans operators use proprietary software to annotate the events. This thesis aims to automate part of the annotation process with a computer vision approach that can recognize one of the most frequent events in soccer: the passes. To achieve this purpose, we combine soccer broadcast videos and events data. Broadcast videos are the input of the models, while the events data define the labels of the videos. We propose a model that is a combination of the pre-trained model ResNet18, applied to extract features from single frames and a Bidirectional LSTM model that analyzes the temporal evolution of the extracted features. Moreover, we use real-time object detection method YOLO to extract the positional information of the ball and the players inside each frame. This information is concatenated to the feature extracted from the ResNet18 model and used as input of bidirectional LSTM. Our results show a significant improvement in the accuracy of pass detection with respect to baseline classifiers applied to the same task, highlighting that our approach is a first step towards the automation of events annotation in soccer.