Review and evaluation of player detection methods in field sports Comparing conventional and deep learning based methods

In order to analyse field sport videos, accurate player detection is essential. If players' positions are identified correctly in video frames, the higher level analysis such as event detection, activity and performance analysis will be more accurate. For player detection, many methods have been presented until now with varying performances and under different evaluation settings. Convolutional Neural Networks (CNN) became very popular for object recognition and detection. Generally, CNN approaches use Gray or RGB image representations for detection. However, other image representations such as shape information image (SIM) and polar transformed shape information image (PSIM) can also be utilized for player detection. Furthermore, recent deep neural networks such as Faster R-CNN, Single Shot Detector (SSD) and Yolo can be applied for player detection using transfer learning. The aim of this study is to investigate and compare the performances of conventional methods, CNNs with different image representations and complex deep learning methods under the same evaluation settings, and on two field hockey datasets. We compare performances for different overlap ratios and for different occlusion cases. This is the first time an extensive evaluation, review and comparison have been conducted for player detection in field sports.

Automated summary

Accurate player detection is crucial for analyzing field sport videos. Various methods, including conventional ones, convolutional neural networks, and deep learning techniques, have been proposed for player detection. This study aims to compare the performances of these methods on field hockey datasets under the same evaluation settings.