Automatic Real-Time Players and Game- Object Detection.

Bulygin, Mikhail

Automatic Real-Time Players and Game- Object Detection.

Files

Bulygin 1039179-2021.pdf (56.89 MB)

Authors

Bulygin, Mikhail

Issue Date

2021-08-01

Language

en

URI

https://theses.ubn.ru.nl/handle/123456789/16090

Abstract

E cient and reliable object detection is the foundation of automatic video broadcasting for sports events. In this project, we together with Spi- ideo investigate the ability of an object detection CNN model to extract the positions of the players and game objects. We use the Transfer Learning approach taking a pre-trained scaled-YOLOv4 model and t it to make pre- dictions on Spiideo's multi-sport data in real-time. Our results show that scaled-YOLOv4 was able to generalize the concept of a game object across 5 sports (football, basketball, handball, hockey, and eld hockey). The model scored the average precision of 0.8266 and recall of 0.4432 for the Game object class across all sports. Besides, we found out that adding data from di erent sports improves the performance of the model for a single sport. Adding multi-sport data improved the performance of a baseline football model by 76% in precision and by 15% in recall for game object detection on the same dataset. Moreover, the model became more robust to variations and changes in surroundings. Our system is able to generate predictions for the new sports that were not included in the training and were never exposed to the model. We conclude our research with a re ection on the results and an outline of possible future research.