Evaluation of DeepLabCut as a Human Markerless Motion Capture Tool

Abstract

There are a variety of motion capture methods available; however, many of them are not well suited for collections outside a laboratory setting. AI markerless motion capture may fit this need, but its implementation and accuracy need to be better understood. Therefore, the purpose of this research was to evaluate the tracking capabilities of DeepLabCut and conditions (complexity of the feature set and camera setup) that can affect its performance. Two markerless networks, a common joint center tracking set and a complex feature set, were trained using 40 participants completing 6 movements that were recorded by 8 cameras. Network retraining and performance evaluation (tested with 10 participants) occurred 3 times for each network. The results from this markerless motion capture research highlight the importance of choosing minimally occluded features of interest and camera positions that maximize the number of frames where the full feature set is visible.