Abstract
Object detection is widely used in the world of sports, its users including training staff,
broadcasters and sports fans. Neural network based classifiers are used together with
other object detection techniques. The aim of this study was to explore the modern
open source based solutions for object detection in sports, in this case for detecting
football players. TensorFlow Object Detection API, an open source framework for
object detection related tasks, was used for training and testing an SSD (Single-Shot
Multibox Detector) with Mobilenet- model. The model was tested as a) pre-trained
and b) with fine-tuning with a dataset consisting of images extracted from video
footage of two football matches. Following hypotheses were examined:
1) Pre-trained model will not work on the data without fine-tuning.
2) Fine-tuned model will work reasonably well on the given data.
3) Fine-tuned model will have problems with occlusion and players pictured
against the rear wall.
4) Using more variable training data will improve results on new images.
The results of this study indicate that:
1) The pre-trained model was useless for detecting players in the test images.
2) A fine-tuned model worked reasonably well.
3) Problem areas were players in clusters and/or pictured against the rear wall.
4) A model trained with data from one game was able to detect players in footage
from another game. The overall model performance did not much improve by
training the model with data from two games.
Other model types (such as Faster R-CNN model) should be tested on the data.
Users
Please
log in to take part in the discussion (add own reviews or comments).