Experimental setup for evaluating the accuracy of markerless human motion estimation techniques
Daniele Bibbo, Michela Goffredo, Silvia Conforto, Maurizio Schmid, Tommaso D’Alessio · Iris (Roma Tre University) · 2004
Human Movement Analysis is generally performed through the exploitation of stereophotogrammetric systems that rely on the use of markers [3]. With these systems, it is possible to reconstruct the trajectories of reflecting objects allocated on specific points of the human body and it is possible to reconstruct limb kinematics through the calibration parameters. In order to reduce measuring errors and artefacts dedicated algorithms are used. The performance of these systems are evaluated through the accuracy of kinematics estimation, using a range of tests of “spot-checks” developed to this purpose. There are two kinds of tests: static and dynamic ones. The first tests are based on the estimation of distances between markers, whose mutual distance is known. The second ones are based on the reconstruction of known trajectories (i.e. rotating disk test, gravity test etc.[7]). An example of static spots -check is the “Movement Analysis Laboratory” test (MAL test [5]): a stiff bar, that presents an extremity laid on a point of the fixed reference system, is used. Two markers are placed on the bar in two different points. The accuracy is evaluated by estimating the coordinates of the extremity as the intersection of the straight line detected by the markers in different positions. An example of dynamic spots -check is the “gravity test”: a rigid pendulum, that presents two markers placed on the extremities, is let be swinging. The accuracy is evaluated comparing the trajectory of the swinging extremity, estimated with the use of markers, with the real one. Marker based systems show relatively high costs while the procedure is generally time consuming and uncomfortable for the patients. In some applications, the use of these specialized systems can be replaced by the utilization of video capture systems that do not necessarily rely on marker application, together with specific image processing techniques for detection and tracking of relevant body points. As it has been done with stereophotogrammetric systems, it is also necessary to define some protocols useful to evaluate the accuracy of markerless systems in the estimation of kinematic parameters. This is possible by using the same rationale of previous “spot-checks” tests concerning stereophotogrammetry, but it is necessary to evaluate specific problems connected with markerless systems and to define an adequate experimental setup. In order to manage this problem, a protocol has been defined and developed considering three significant phases: 1) Establishing kinematic features of a mobile reference system through a direct measure in order to have a valid term of comparison. 2) Estimating kinematic features of the mobile reference system through image processing techniques, using dedicated algorithms. 3) Comparing kinematic features through statistical analysis choosing appropriate evaluation parameters. This work aims at defining a valid experimental setup to evaluate the accuracy of two markerless techniques when used to estimate motion from videos: a Block Matching Algorithm and a Transformed Domain approach.