Using Markerless Motion Capture and Principal Component Analysis to Classify BMX Freestyle Tricks

Abstract

This thesis presents a novel Bicycle Motocross (BMX) Freestyle (FS) trick classification technique developed for the Australian Cycling Team. The first step is tracking six key points on the athlete and their bike using DeepLabCut, an opensource markerless motion capture software. Next, a Principal Component Analysis (PCA) is applied to the tracking data to calculate metrics to identify each trick type. Finally, a classifier is trained to learn these metrics. The dataset used in this paper focused on three common BMX Freestyle tricks: 360, backflip, and flair. The Logistic Regression model achieved the highest accuracy among the classifiers, correctly predicting the trick for 94.2% of the instances. This thesis discusses other ways to apply this data, such as novel trick generation. It also examines the robustness and cost benefit trade off of the classifier.