Comparison of Visual Trackers for Biomechanical Analysis of Running

TL;DR

This study assesses six pose-tracking methods—two point trackers and four joint trackers—for biomechanical analysis of sprinting, using the VideoRun2D dataset with ground-truth annotations from biomechanics experts. It introduces a post-processing module for outlier detection and smoothing and demonstrates that joint-based trackers, especially when combined with post-processing, achieve substantial accuracy improvements (RMSE around $8^\circ$ for hip and $5$–$6^\circ$ for knee) compared with ground truth. A late-fusion approach across top trackers further reduces error on occluded left-side joints, illustrating the value of ensemble predictions in markerless sprint biomechanics. The findings support the utility of markerless pose tracking for practical biomechanical analysis while highlighting remaining gaps for high-precision applications and suggesting directions for future enhancement through model fusion and multi-view recording.

Abstract

Human pose estimation has witnessed significant advancements in recent years, mainly due to the integration of deep learning models, the availability of a vast amount of data, and large computational resources. These developments have led to highly accurate body tracking systems, which have direct applications in sports analysis and performance evaluation. This work analyzes the performance of six trackers: two point trackers and four joint trackers for biomechanical analysis in sprints. The proposed framework compares the results obtained from these pose trackers with the manual annotations of biomechanical experts for more than 5870 frames. The experimental framework employs forty sprints from five professional runners, focusing on three key angles in sprint biomechanics: trunk inclination, hip flex extension, and knee flex extension. We propose a post-processing module for outlier detection and fusion prediction in the joint angles. The experimental results demonstrate that using joint-based models yields root mean squared errors ranging from 11.41° to 4.37°. When integrated with the post-processing modules, these errors can be reduced to 6.99° and 3.88°, respectively. The experimental findings suggest that human pose tracking approaches can be valuable resources for the biomechanical analysis of running. However, there is still room for improvement in applications where high accuracy is required.

Figures (6)

• Figure 1: Data acquisition sketch and example of the different stages of a stride for an individual in the dataset.
• Figure 2: Location of the joint points marked on Kinovea software (ground truth).
• Figure 3: Block diagram of the VideoRun2D system. The system has five modules that estimate joint angles.
• Figure 4: Graphical example of the correction of the left knee angle in all the thirteen strides from the sprinter two, both before (on the left) and after (on the right) post-processing
• Figure 5: Graphical representation of joint points and angles calculated for four sprinters using the RTMPose model after the proposed tracking post-processing. The orange lines and dots represent the right side of the corridor, while the purple lines and dots represent the left side. (Color image.)