Reconstruction of skeletal movement using skin markers: comparative assessment of bone pose estimators
© Cereatti et al; licensee BioMed Central Ltd. 2006
Received: 01 December 2005
Accepted: 23 March 2006
Published: 23 March 2006
The assessment of the accuracy of the pose estimation of human bones and consequent joint kinematics is of primary relevance in human movement analysis. This study evaluated the performance of selected pose estimators in reducing the effects of instrumental errors, soft tissue artifacts and anatomical landmark mislocations occurring at the thigh on the determination of the knee kinematics.
The pattern of a typical knee flexion-extension during a gait cycle was fed into a knee model which generated a six-components knee kinematics and relevant marker trajectories. The marker trajectories were corrupted with both instrumental noise and soft tissue artifacts. Two different cluster configurations (4 and 12-marker cluster) were investigated. Four selected pose estimators, a Geometrical method, a SVD-based method, and the Pointer Cluster Technique in the optimized and non optimized version, were analyzed. The estimated knee kinematics were compared to the nominal kinematics in order to evaluate the accuracy of the selected pose estimators.
Results have shown that optimal pose estimators perform better than traditional geometric pose estimators when soft tissue artifacts are present. The use of redundant markers improved in some cases the estimation of the dynamics of the kinematics patterns, while it does not reduce the offsets from the nominal kinematics curves. Overall, the best performance was obtained by the SVD-based pose estimator, while the performance of the PCT pose estimator in its optimal version was not satisfactory. However, the knee kinematics errors reached 5 deg for rotations and 10 mm for translations).
Given the favorable experimental conditions of this study (soft tissue artifacts determined from a young, healthy and non overweight subject), the errors found in estimating the knee kinematics have to be considered unsatisfactory even if the best performing pose estimator is used. Therefore, it is the authors' opinion that the movement analysis research community should make additional efforts in the search of more subject specific error models to increase the accuracy of joint kinematics estimations.
In the last two decades, numerous human movement scientists faced the problem of the in vivo and non-invasive reconstruction of the skeletal movement from measures of the 3D position of points located on the human body surface. To this purpose, the estimation of the pose (position and orientation) of skeletal bones is required. In order to determine the pose of a bone, the definition of a reference frame (RF) attached to it, typically an orthogonal RF defined by bone anatomical landmarks (ALs), is required (anatomical RF). In order to gain more freedom in positioning the markers on the body segment, a technical RF is often identified from the position of the markers attached to the body segment. In this case, the rigid body transformation parameters between the two systems of axes can be obtained with a calibration procedure of the bone ALs .
Three sources of errors typically affect the in vivo estimation of the pose of a skeletal bone when non invasive techniques are used: instrumental errors, soft tissues artifacts and AL mislocation. Reviews of such sources of errors and their effect on joint kinematics were recently published [2–5]. Numerous techniques have been proposed to reduce the effect of one or more of these sources of error. Some of them were based on marker attachment devices , others were based on computational methods. The latest motion capture technology allows the positioning of a high number of markers on all sides of the surface of each body segment, thus facilitating the use of redundant markers attached directly to the subject skin, which renders in some cases the use of attachment devices obsolete. The same consensus has not been reached yet in the determination of an "optimal" computational method for the determination of the bone pose when the above mentioned sources of error are present.
Skeletal bone pose has been determined from marker positions using various methods including geometrical and optimal. Usually the geometrical pose estimators do not make use of redundant information included in the marker position. Typically, an axis of the RF is defined as the oriented line passing through two of the markers attached to the body segment, a second is defined as the axis perpendicular to the first and lying on the plane identified by the three markers, and the third one is obtained from the cross product of the unit vectors of the two axes already defined. Anatomical RFs are normally identified by the relevant AL positions using geometrical methods with few exceptions . The optimal pose estimators derive the rigid body transformation parameters using least squares methods to solve the difference equation by using either the matrix characteristic equation , the singular value decomposition (SVD) [9–13] or iterative methods . In some cases, non-least squares method are used .
All the above mentioned pose estimators have pros and cons the importance of which is difficult to determine with in vivo experiments without using invasive techniques. In fact, while the estimation of instrumental errors is not problematic, the in vivo quantification of soft tissue artifacts and their effect on the determination of the bone pose is still an open issue. Recently, several attempts have been made to this respect using various approaches [16, 17]. Simulation studies have also been proposed with the limitation of imposing error time histories not generated from experimental observations .
Some studies compared the pose estimators mentioned above [13, 19]. In the latter study the goal was to test the examined pose estimators' performance in the case of ill-conditioned marker distribution, since some of the optimal pose estimators are very sensitive to highly symmetric spatial distribution of the marker cluster.
Some of the mentioned studies focused their attention on a specific joint, the knee, or a specific segment, the thigh, since soft tissues artifacts have great effects on the determination of the thigh kinematics, which is of particular relevance in gait analysis.
In this study, we evaluated the performance of various pose estimators by implementing a four-bar linkage (FBL) model of the knee attached to rigid models of the tibia end the femur and relevant ALs. The latter bone was equipped with virtual markers the trajectories of which, during simulated FBL movement, were corrupted by both instrumental noise and soft tissue artifacts. By including AL mislocation information in the model, an evaluation of the performance of the selected pose estimators was obtained.
The knee model
Parameters of the four bar linkage model of the knee.
FBL parameters [mm]
knee parameters [deg]
The anthropometric and kinematic data
Generation of noisy movement data
Mean and standard deviation values of the soft tissue artifacts expressed in the femoral anatomical frame applied to the nominal positions of the twelve thigh markers. Data regarding the markers used in the 4-marker configuration are in bold. X positive direction: anterior; Y positive direction: proximal; Z positive direction: from left to right.
In order to quantify the combined influence on the estimated kinematics of instrumental noise, soft tissue artifacts and AL mislocation, the medio-lateral axis of the femur anatomical RF (the same used by Grood and Suntay  for representing knee flexion-extension) was made to rotate ± 5 deg about the femur long axis.
Bone pose estimators
Once the two sets of thigh marker trajectories containing (a) instrumental noise and (b) instrumental noise added to soft tissue artifacts were generated, four different bone pose estimators were applied to them so that the pose of a technical RF could be determined. Thus, the anatomical RF pose could be estimated using the information obtained by the AL calibration. Finally, the knee model kinematics could be determined using the method proposed by Grood and Suntay  and compared to the nominal knee kinematics.
The following four bone pose estimators were tested in this study.
A geometrical method using the selected four markers. The origin of the technical frame was located at the midpoint of the line joining ATH2 and LTH3. The direction of the line joining the origin and ATH1 pointing at ATH1 was projected on the plane formed by ATH2, LTH3 and PTH1 defined the Y axis. The axis perpendicular to the Y-axis and lying on the plane formed by ATH2, LTH3 and PTH1 pointing to the right was used as the Z-axis. The X-axis was obtained from the cross product of the unit vectors of the Y-axis and the Z-axis.
The SVD-based pose estimator. The SVD-based pose estimator finds the pose of a reference marker configuration (typically obtained in static conditions) that minimizes the elastic energy of identical virtual springs connecting the corrupted marker locations (recorded during the movement) to the relevant marker location in the reference configuration. The SVD-based pose estimator was implemented as in Cappozzo et al. .
The Pointer Cluster Technique (PCT) based on the identification of the marker cluster tensor of inertia . The PCT pose estimator consists of two steps: first, a single mass value (typically a unit value) is assigned to all markers forming the cluster attached to a body segment. In every instant i during the movement the tensor of inertia of the configuration D i of the corrupted marker locations is generated and the pose of the RF formed by the principal axes of inertia of D i is determined (PCTu). Secondly, the marker mass values are adjusted until the difference between the sum Λ i of the square values of the tensor of inertia eigenvalues obtained from the marker cluster during movement and the sum Λ0 resulting from a reference marker configuration D0, is minimum (PCTa). Both steps were implemented. The PCTa objective function was minimized using the function "lsqnonlin" (Gauss-Newton method) available in Matlab (MathWorks inc.). As suggested by the authors, the termination tolerance of the function value was set equal to 0.1. Occasionally, both PCTu and PCTa pose estimators produced instantaneous orientation inversions of the principal axes that were removed by comparing axes orientation in adjacent time instants.
Results and discussion
The 12-marker configuration prevented the use of a geometric estimator and allowed a slight improvement of the knee kinematics estimation (PCTu and SVD-based), more evident in the two minor rotational components (ab-adduction and internal-external rotation). However, since the difference between the kinematic patterns obtained from each of the various pose estimators and the nominal kinematics were limited for both rotations and translations (< 2 deg and < 4 mm, respectively), the effects of the instrumental noise could be considered as less important. Therefore, when soft tissue artifacts are absent (as in the case of the gait analysis of subjects with prostheses and/or orthoses), in most cases it is not necessary to use more than three or four markers and a geometrical pose estimator can be satisfactory. Knee kinematics PCTa estimates presented irregular patterns. This undesired effect is due to the fact that in each instant of time the optimization function used in the PCTa, instead of operating on each eigenvalue of the tensor of inertia of the marker configuration, minimizes the difference between Λ and Λ0, the sums of the squares of the three eigenvalues. As a consequence, the proportions among the eigenvalues values can vary (and therefore the relevant eigenvectors) keeping the sum Λ very close to Λ0. The frame of the plots in Figure 4b marked with a star (*) is an example of such malfunctioning. The eigenvalues of the reference configuration D0 were 50190 mm2, 81340 mm2 and 113790 mm2, those of the instantaneous tensor of inertia of the configuration D1 minimizing the optimization function at the first frame were 48690 mm2, 82380 mm2 and 113700 mm2, and those obtained at the selected frame (D*) were 40410 mm2, 87450 mm2 and 113150 mm2, respectively. Whereas the eigenvalues of D1 were close to those of D0, the eigenvalues of D* were remarkably different and so were the relevant eigenvectors. Therefore, the axes of inertia of the marker configuration at the selected frame were different from the reference ones. Such difference reflected on the low accuracy of the knee kinematics estimate.
Mean and standard deviation values of the error patterns of the knee kinematics reported in Figure 5.
Mean and standard deviation values of the error patterns reported in Figure 6.
SVD (+5 deg)
PCTu (+5 deg)
SVD (-5 deg)
PCTu (-5 deg)
The study focused on the knee kinematics estimation from noisy thigh data. The shank pose estimation was considered noise-free for two reasons: a) results interpretation is simpler when the joint kinematics error contributions are presented separately for the two segments forming the joint and b) the errors affecting the thigh pose estimation are typically higher than those affecting the shank pose estimation (larger thigh soft tissue artifacts and higher risk of thigh AL mislocation).
The soft tissue artifacts patterns used in this study were obtained from a young healthy, non overweight subject. For subjects with different characteristics, soft tissue artifacts could be more disruptive, therefore the errors shown in this study may represent an estimate of a minimum possible error.
The present study was carried out to report the level of accuracy that can be reached in estimating bone pose and joint kinematics by some bone pose estimator methods selected from those proposed in the last two decades. The study took into account all sources of errors typically affecting joint kinematics estimation. By focusing on the knee kinematics estimated from corrupted thigh marker trajectories, this study highlighted the advantages and the limitations of both simple and sophisticated pose estimators. Overall, the SVD-based pose estimator showed the best performance. However, errors up to 5 degrees for rotations and 10 mm for translations were found even in the most favorable conditions. Therefore, since a large inter-subject variability of thigh soft tissue artifacts is commonly recognized, it is the authors' opinion that future research should aim at using more subject specific data to correct the effects of both the soft tissue artifacts and AL mislocations.
- Cappozzo A, Catani F, Della Croce U, Leardini A: Position and orientation in space of bones during movement: anatomical frame definition and determination. Clin Biomech (Bristol, Avon) 1995, 10: 171-178. 10.1016/0268-0033(95)91394-TView ArticleGoogle Scholar
- Cappozzo A, Della Croce U, Leardini A, Chiari L: Human movement analysis using stereophotogrammetry. Part 1: theoretical background. Gait Posture 2005, 21: 186-196.PubMedGoogle Scholar
- Chiari L, Della Croce U, Leardini A, Cappozzo A: Human movement analysis using stereophotogrammetry. Part 2: instrumental errors. Gait Posture 2005, 21: 197-211. 10.1016/j.gaitpost.2004.04.004View ArticlePubMedGoogle Scholar
- Leardini A, Chiari L, Della Croce U, Cappozzo A: Human movement analysis using stereophotogrammetry. Part 3. Soft tissue artifact assessment and compensation. Gait Posture 2005, 21: 212-225. 10.1016/j.gaitpost.2004.05.002View ArticlePubMedGoogle Scholar
- Della Croce U, Leardini A, Chiari L, Cappozzo A: Human movement analysis using stereophotogrammetry. Part 4: assessment of anatomical landmark misplacement and its effects on joint kinematics. Gait Posture 2005, 21: 226-237. 10.1016/j.gaitpost.2004.05.003View ArticlePubMedGoogle Scholar
- Cappozzo A, Catani F, Leardini A, Benedetti MG, Della Croce U: Position and orientation in space of bones during movement: experimental artefacts. Clin Biomech (Bristol, Avon) 1996, 11: 90-100. 10.1016/0268-0033(95)00046-1View ArticleGoogle Scholar
- Della Croce U, Camomilla V, Leardini A, Cappozzo A: Femoral anatomical frame: assessment of various definitions. Med Eng Phys 2003, 25: 425-431. 10.1016/S1350-4533(03)00032-8View ArticlePubMedGoogle Scholar
- Veldpaus FE, Woltring HJ, Dortmans LJ: A least-squares algorithm for the equiform transformation from spatial marker co-ordinates. Journal of Biomechanics 1988, 21: 45-54. 10.1016/0021-9290(88)90190-XView ArticlePubMedGoogle Scholar
- Arun KS, Huang TS, Blostein SD: Least-Squares Fitting of 2 3-D Point Sets. IEEE Trans Pattern Anal Mach Intell 1987, 9: 699-700.Google Scholar
- Soderkvist I, Wedin PA: Determining the movements of the skeleton using well-configured markers. Journal of Biomechanics 1993, 26: 1473-1477. 10.1016/0021-9290(93)90098-YView ArticlePubMedGoogle Scholar
- Challis JH: An examination of procedures for determining body segment attitude and position from noisy biomechanical data. Medical Engineering & Physics 1995, 17: 83-90. 10.1016/1350-4533(95)91877-JView ArticleGoogle Scholar
- Cappozzo A, Cappello A, Della Croce U, Pensalfini F: Surface-marker cluster design criteria for 3-D bone movement reconstruction. IEEE Trans Biomed Eng 1997, 44: 1165-1174. 10.1109/10.649988View ArticlePubMedGoogle Scholar
- Cheze L, Fregly BJ, Dimnet J: A solidification procedure to facilitate kinematic analyses based on video system data. Journal of Biomechanics 1995, 28: 879-884. 10.1016/0021-9290(95)95278-DView ArticlePubMedGoogle Scholar
- Cappello A, La Palombara PF, Leardini A: Optimization and smoothing techniques in movement analysis. Int J Bio-Med Comput 1996, 41: 137-151. 10.1016/0020-7101(96)01167-1View ArticleGoogle Scholar
- Andriacchi TP, Alexander EJ, Toney MK, Dyrby C, Sum J: A point cluster method for in vivo motion analysis: applied to a study of knee kinematics. Journal of Biomechanical Engineering 1998, 120: 743-749. 10.1115/1.2834888View ArticlePubMedGoogle Scholar
- Camomilla V, Donati M, Cappozzo A: Non-invasive soft tissue artefact assessment: September 5-7th 2005; University of Salford, Salford. ; 2005:80-81.Google Scholar
- Cappello A, Stagni R, Fantozzi S, Leardini A: Soft tissue artifact compensation in knee kinematics by double anatomical landmark calibration: performance of a novel method during selected motor tasks. IEEE Trans Biomed Eng 2005, 52: 992-998. 10.1109/TBME.2005.846728View ArticlePubMedGoogle Scholar
- Alexander EJ, Andriacchi TP: Correcting for deformation in skin-based marker systems. Journal of Biomechanics 2001, 34: 355-361. 10.1016/S0021-9290(00)00192-5View ArticlePubMedGoogle Scholar
- Carman AB, Milburn PD: Determining rigid body transformation parameters from ill-conditioned spatial marker co-ordinates. Journal of Biomechanics 2005., In Press, available online:Google Scholar
- Gill HS, O'Connor JJ: Biarticulating two-dimensional computer model of the human patellofemoral joint. Clinical Biomechanics 1996, 11: 81-89. 10.1016/0268-0033(95)00021-6View ArticlePubMedGoogle Scholar
- Camomilla V, Cereatti A, Vannozzi G, Cappozzo A: An optimized protocol for hip joint centre determination using the functional method. J Biomech 2005, 39: 1096-106. 10.1016/j.jbiomech.2005.02.008View ArticleGoogle Scholar
- Grood ES, Suntay WJ: A joint coordinate system for the clinical description of three-dimensional motions: application to the knee. Journal of Biomechanical Engineering 1983, 105: 136-144.View ArticlePubMedGoogle Scholar
This article is published under license to BioMed Central Ltd. This is an Open Access article distributed under the terms of the Creative Commons Attribution License (http://creativecommons.org/licenses/by/2.0), which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.