3D Estimation of Joints for Motion Analysis in Sports Medicine : A study examining the possibility for monocular 3D estimation to be used as motion analysis for applications within sports with the goal to prevent injury and improve sport specific motion

Abstract: 3D joint estimation can be used to track bodies in areas such as entertainment, sports, biomedicine and surveillance to identify bodies from video streams and images. This is most commonly done with multi-view solutions but researchers are currently spending a large amount of resources into developing mono-view solutions. The idea is to utilise neural networks to identify 3D joints by exploiting patterns and restrictions found in the human pose. Currently these systems are showing great results in controlled settings with good accuracy. However, for this to become a widespread technique it will be crucial for the systems to be able to perform with high accuracy in all types of settings. This thesis will focus on evaluating if current systems could be used to perform 3D estimations with high accuracy on movement analysis in sports settings. Based on a prestudy performing meaningful analysis in the area would require the system to perform with an accuracy of 4 cm. In order to evaluate the accuracy in this setting this thesis consists of three steps. Firstly two methods are picked by performing a prestudy of currently available monocular 3D joint estimation solutions. The accuracy of these two methods is then evaluated on two datasets, one which both have been trained on and another sports focused dataset which neither have been trained on. The sports dataset consisted of video sequences of movements from tennis, volleyball, basketball, badminton, football and rugby. In the last step a smoothing filter is applied on the results from the method that performed best on the sports dataset. This was done in order to further improve the accuracy of the system and evaluate the idea of using such techniques. The comparisons were made by measuring the mean per joint positional error for each of the images in the datasets. Both of the methods used were unable to reach a 4 cm accuracy on the sports dataset and thus are not suitable for this type of analysis in their current states. However applying a filter on the results did result in a small improvement of the accuracy and could be an area of research to look further into when these methods are further developed.