Abstract
AbstractBackgroundSkeleton tracking is a valuable tool for monitoring user performance and has been performed with different sensors and software tools. It enables to quantify 3D motion accurately. We present a comparative analysis of cameras and software tools utilized in skeleton tracking. We describe an experimental design and analysis methods with four steps. Specifically, we conducted an analysis comparing two ZED-2i cameras with a RealSense camera using commercial skeleton tracking algorithms.MethodsThe comparative analysis methodology describes the experimental design and analysis methods for comparing kinematic data from both upper and lower extremities. An experiment to evaluate the performance of three different cameras (RealSense, ZED2mm, and ZED4mm) and three skeleton tracking algorithms (PyZED, Nuitrack, and MediaPipe) was performed with 16 participants. The participants were required to walk along a 6-meter-long path while moving their hands up and down horizontally. Ground truth was measured by a Vicon 3D motion analysis system. Two groups of features corresponding to the upper and lower extremities were derived from the raw data and compared using two methods. In the first method, the RMSE values were compared to the data from the Vicon system. The RMSE results were also combined using two grading techniques with different feature importance weights to determine the camera with the lowest RMSE. In the second method, a linear mix model with a Wald statistical test was applied.ResultsThe ZED-2i cameras outperformed the RealSense camera across upper- and lower-extremity features and different weights. The ZED2mm camera showed better average median grading, lower standard deviation, and superior performance in the model analysis.ConclusionThe comparative analysis methodology provides a systematic method to compare software tools and cameras for skeleton tracking. The ranking enables to prioritize different features. Skeleton tracking with the ZED cameras yielded better performances for both upper and lower extremities features.
Publisher
Cold Spring Harbor Laboratory