Poster Session II
Evaluation of a novel approach to generating three-dimensional data using markerless tracking
Project Type
Poster
Project Funding and Affiliations
Support from Immergo Labs
Faculty Mentor’s Full Name
Matthew W Bundle
Faculty Mentor’s Department
Integrative Physiology and Athletic Training
Abstract / Artist's Statement
Classical methods to analyze human movement trajectories, typically require a system of cameras to capture the location of anatomical landmarks highlighted by reflective markers. If the equipment is pre-calibrated the per-camera 2-dimensional position data can resolve the 3D location of a joint or body segment. To ensure 1mm levels of accuracy, these systems are labor intensive, suffer from a lack of portability, and are difficult to deploy beyond the laboratory or clinic. Recent techniques of automated image analysis, convert traditional images into a heatmap that locates the human body and then relies on additional learning and contextual information to predict segment locations. This analysis does not require markers, can be fully automated, occur in real-time, and be deployed anywhere. Nonetheless, previous work has found the 3D data obtained with markerless tracking to have low accuracy and high inter-trial variability(Horsak et al. 2024). Based on our laboratory’s experience, we hypothesize the accuracy loss is due to the unconventional reconstruction of the 3D movements rather than the 2D pose estimation. We developed an analysis workflow using the classical direct linear transformation, to evaluate whether 2D pose estimates can reconstruct 3D movement trajectories with an accuracy approaching laboratory equipment. The workflow consists of a wand calibration to determine the DLT coefficients, an open-source pose estimator(RTMpose), a custom signal conditioning algorithm, and the academically validated motion analysis platform DLTdv(Hedrick 2008) for the 3D transform. Seven subjects were simultaneously assessed using our approach and a laboratory grade system to determine the accuracy of the technique.
Category
Physical Sciences
Evaluation of a novel approach to generating three-dimensional data using markerless tracking
UC South Ballroom
Classical methods to analyze human movement trajectories, typically require a system of cameras to capture the location of anatomical landmarks highlighted by reflective markers. If the equipment is pre-calibrated the per-camera 2-dimensional position data can resolve the 3D location of a joint or body segment. To ensure 1mm levels of accuracy, these systems are labor intensive, suffer from a lack of portability, and are difficult to deploy beyond the laboratory or clinic. Recent techniques of automated image analysis, convert traditional images into a heatmap that locates the human body and then relies on additional learning and contextual information to predict segment locations. This analysis does not require markers, can be fully automated, occur in real-time, and be deployed anywhere. Nonetheless, previous work has found the 3D data obtained with markerless tracking to have low accuracy and high inter-trial variability(Horsak et al. 2024). Based on our laboratory’s experience, we hypothesize the accuracy loss is due to the unconventional reconstruction of the 3D movements rather than the 2D pose estimation. We developed an analysis workflow using the classical direct linear transformation, to evaluate whether 2D pose estimates can reconstruct 3D movement trajectories with an accuracy approaching laboratory equipment. The workflow consists of a wand calibration to determine the DLT coefficients, an open-source pose estimator(RTMpose), a custom signal conditioning algorithm, and the academically validated motion analysis platform DLTdv(Hedrick 2008) for the 3D transform. Seven subjects were simultaneously assessed using our approach and a laboratory grade system to determine the accuracy of the technique.