The Use of Physical Props in Motion Capture Studies

Transcription

1 Copyright 2008 SAE International 08DHM-0049 The Use of Physical Props in Motion Capture Studies Monica L. H. Jones, Jim Chiang and Allison Stephens Ford Motor Company, Michigan, USA Jim R. Potvin McMaster University, Ontario, CANADA ABSTRACT It is generally accepted that all postures obtained from motion capture technology are realistic and accurate. Physical props are used to enable a subject to interact more realistically within a given virtual environment, yet, there is little data or guidance in the literature characterizing the use of such physical props in motion capture studies and how these effect the accuracy of postures captured. This study was designed to evaluate the effects of various levels of physical prop complexity on the motion-capture of a wide variety of automotive assembly tasks. Twenty-three subjects participated in the study, completing twelve common assembly tasks which were mocked up in a lab environment. There were 3 separate conditions of physical props: Crude, Buck, and Real. The Crude condition provided very basic props, or no props at all, while the Buck condition was a more elaborate attempt to provide detailed props. Lastly, the Real condition included real vehicle sections and real parts. Plant operator subjects were also provided video feedback of themselves performing the task in the actual assembly plant, to ensure a similar method was used in the lab. Analyses of the postures adopted for each of prop levels indicate that there are differences associated with each. Relative to the motion capture using real automotive parts, the differences decreased as the assessment progressed from crude to more complicated physical props. INTRODUCTION Virtual reality (VR) and motion capture (MOCAP) technology provide significant potential for obtaining accurate whole-body movements (Wilson, 1999). The postures captured in a MOCAP simulation are based upon the movements of real subjects. Specifically, this technology allows an operator to become immersed within a virtual environment. Body markers are used to track and transpose the human subject onto a digital human model (DHM) manikin. These movements can then be analyzed using a software package such as Jack (Siemens PLM Software Inc, Plano, TX) as the DHM manikin provides all the inputs necessary for a comprehensive ergonomic assessment. This approach can improve the simulation time, provide more accurate posture predictions and capture the variability in movement represented on the assembly plant floor. It is generally accepted that all postures obtained from motion capture technology are realistic and "assembly representative". Yet, caution must be emphasized in that there is a broad need for direct comparisons of human behaviour in virtual and real environments (Stoffregen at al., 2003). To enhance the motion capture process, video of an assembly operation in question is often provided to parameterize the industrial workstation. Further, the virtual computer-aided (CAD) environment should be built to replicate the assembly facility. The overall objective is for the subject to interact with both the virtual and physical environment in a motion capture lab in order to identify critical movements of the assembly operation in question. Physical props are also used to enable a subject to interact more realistically within a given virtual environment. While complete physical replication in the virtual environment would yield the most realistic results, this would be both time consuming and costly. It is proposed that the degree to which physical props are necessary is dependent upon how the subject will interact with the virtual environment. For example, if a simulated posture involves bracing or leaning relative to geometric CAD data, it is reasonable to assume that an appropriate physical hard contact point is required for the subject (Figure 1). The process of determining the type, complexity and number of physical props required for an assembly representative motion capture simulation is largely undefined. The objective of the current study is to evaluate the postural differences that result under different levels of physical props compared to Real operators performing the task with real parts and vehicle sections from the assembly facility using motion capture technology. The goal of this study to help define the degree of physical props complexity and detail that is required during a motion capture study for each of the assembly tasks studied.

2 Subjects Seventeen male and six female subjects, ranging from years of age, participated in this study. Experience level and training of the individual subjects also ranged from experienced assembly operators (n=12) to non-operators/engineers (n=11). Summary demographics are presented in Table 1. For each workstation, a total of 3 assembly operators had experience with the task and were defined as the Real subject. For example, for Task #3, subjects 7, 8 and 9 served as the Real subjects and they were subsequently tested under Crude and Buck conditions for the other 11 tasks (for which they had no previous work assembly experience). Table 1: Mean summary of subject descriptors Figure 1: The images above show the physical props (Crude or Buck Level) (top) used to replicate a virtual environment (bottom) while evaluating the installation of a foam piece at the centerline of the vehicle. METHODS This study focused on understanding the level of physical propping that is necessary to capture a posture and accurately reflect the operation in assembly production. To this end, virtual and physical environments were constructed for 12 assembly tasks at varying degrees of complexity (Figure 1). The tasks had varying levels of prop complexity and subjects would use the props to interact more realistically with the selected environment. Visual feedback, of the subject s interactions with the virtual CAD geometry and workstation environment, was provided in all trials. The purpose of this design was to progress upwards in physical prop complexity from Crude to Buck props so that the joint kinematic variables could be compared with that from the Real subjects. This will help determine how complex and precise physical prop levels need to be to achieve valid and reliable results that compare favourable with what would have been obtained from Real workers in the manufacturing environment. Plant Workstation Subject Sex Ht (m) Mass (kg) 1 M M Car 3 F Plant 4 M M M M M Trunk 9 M Plant 10 M M M Pilot Plant 13 F F M M F Engineer 18 F M F M M M Average DATA COLLECTED The study was conducted in the Human Simulation Laboratory at Ford Motor Company. Data were collected using a laboratory set-up comprising of a motion capture system, visual display of the virtual workstation and physical props. A Motion Analysis Corporation (Santa Rose, CA) passive optical system was used with 18 Hawk cameras collecting at a rate of 60 fps. Fifty-two markers were placed on each subject. A Calcium Solver skeleton was subsequently fit to the marker locations in the Motion Analysis EvaRT software. A projector displayed a real time image of the subject performing the task in the virtual environment on one of three walls for visual feedback. Tasks There were 12 individual assembly tasks in this study, 6 from a car assembly plant and 6 from a truck assembly plant. Before collection in the virtual lab, two operations from each of the assembly plants were evaluated for two individual operators and each were video recorded for

3 approximately 5 job cycles. The operations were further broken down into functional tasks. A functional task was defined by the manipulation of a specific part or the use of a tool. Each operation included approximately 3 subtasks (n = 2*2*3 =12). See the Appendix for a pictorial description of the 12 tasks. Apart from the workstations being selected based on vehicle type, they were also selected based on work zone (i.e. the horizontal and vertical reach that the operator would have to apply the force in order to complete the job). The installation efforts for each task were derived from existing surrogate data or handheld force gauge testing on a limited sample size. Physical Prop Levels There were three levels of physical props evaluated. The Crude level required subjects to perform the assembly task with minimal physical props. It provided very basic props or no props at all (i.e. only the virtual environment conveying the relationship of the subject with respect to the vehicle and hand load location). The Buck condition involved a more elaborate attempt to provide detailed physical materials, including a mix of basic props (such as a lift table, resizable door openings, etc). A fundamental characteristic of the Buck condition was to provide feedback of physical hardpoints to the subjects. Figure 2 illustrates the difference between the Crude and Buck prop levels for task #10. The final method, entitled Real, included real vehicle sections and real parts. Real subjects were also provided with video feedback of themselves or other Real subjects performing the task in the actual assembly plant, to ensure a similar method was used in the lab. This was the best lab representation of the exact task performance and was repeated and simulated in the motion capture lab so that the body postures could be recorded for analysis. However, in some cases there was little difference between conditions (e.g. for Tasks 10-12, both the Buck and Real conditions were based on the vehicle section. a b Figure 2: Physical Prop Conditions: a) CRUDE level Task #10; b) BUCK level Task #10 PROTOCOL The three Prop Levels (Crude, Buck and Real) and 12 Tasks were presented in a randomized order within each of the four workstations. For tasks that subjects performed as Real, they were not tested with the Crude and Buck props. Each subject was given a set of instructions describing what they were required to do for each task. For each Task/Method combination, each subject would perform a practice session and then 5 trials were collected through motion capture. Each trial included: 1) the subject standing in a T-pose to ensure proper marker identification (arms abducted 90 deg to the side), 2) the subject performing the Task and holding the end position for 3 seconds, 3) a repeat of the T-pose. DATA ANALYSIS Three out of the five trials from each subject were processed for each Task/Method combination. Generally, the last 3 trials were used. However, if any of those trials had incomplete data, one of the first two trials was used as its replacement. The threedimensional motion capture marker data was preprocessed in EvaRT software and a skeleton was derived. The skeleton was a linkage representation of the human body in the terminal posture which was mapped to a manikin in Jack software (Siemens PLM Software Inc, Plano, TX). The manikin used in Jack was based on the segment lengths from each individual subject and was the same sex as the subject. Dependent Variables Some key joint angle variables were recorded from the Jack outputs and these included all angles from the elbows, shoulder and trunk. Given that some tasks were performed with one hand, and this hand differed between subjects, weighted joint averages were also calculated. Weighted joint averages combined the right and left sides with values weighted by how much force was applied with that side. For example, when only the right hand was used, then the weighted elbow angle would be the same as the right elbow angle. When the right hand was used with 20 N and the left hand was used with 30 N, then the weighted elbow angle would be based on a 40% weighting of the right elbow angle and a 60% weighting of the left elbow angle. This calculation served to decrease the between-subject variability caused by the wide variety of postures observed for the unloaded arms, especially when subjects differed in which hand they used for one handed tasks. Various marker-to-marker distances were also calculated. For any loaded hand, the moment arm was estimated from the hand to both the shoulder and L5/S1. For anterior/posterior pushes and pulls (Tasks #5 & 8), this was calculated as the height difference between the hand and shoulder or L5/S1 (along the vertical Y axis). For vertical loads (Tasks #1, 3, 4, 6, 7, 10, 11) this was calculated as the horizontal distance from the hand to either joint (on the XZ plane). For lateral pushes (Tasks #2, 9, 12) this was calculated as the 3D reach distance from the shoulder to either joint (XYZ resultant).

4 RESULTS Weighted Elbow Flexion Across all of the tasks the Real subjects adopted postures with the highest degree of weighted elbow flexion, with an average of 42 degrees, while the Crude and Buck physical prop levels both averaged 35 degrees of elbow flexion (Figure 3). The RMS errors with Real were approximately 15 deg, while the differences between Crude and Buck were only 9 deg (Table 2). Figure 4: Interaction of Task and Method for resultant shoulder deviation angle (combining abduction/adduction and forward/backward rotation weighted to the hand load). Trunk Flexion Figure 3: Interaction of Task and Method for weighted elbow flexion angle. Weighted Resultant Shoulder Rotation All methods were usually found to have similar amounts of shoulder forward/backward rotation for each task. When pooled across tasks, these rotations were 66 degrees for Real subjects and 64 degrees for both Crude and Buck levels. Larger differences were observed for the shoulder abduction-adduction rotations. The Real subjects averaged 85 degree angle, while the other physical prop conditions were associated with a higher degree abduction-adduction angle (~98 degrees). This difference translated to the weighted resultant shoulder rotation which averaged a degree of rotation collapsed across all tasks for the Real subjects and 120 degree for both the physical prop levels (Figure 4). The differences with Real and between Crude and Buck were similar to those discussed above for elbow flexion (Table 2). Averages of trunk flexion, lateral bend and axial rotation angles are plotted below (Figure 5). In 9 of the 12 tasks, the Real subjects were observed to maintain upright trunk posture, ranging from -10 degrees (extension) to +5 degrees (flexion). Pooled across the 12 tasks, the Crude and Buck levels physical prop levels were observed to have more degrees of flexion than Real. The pooled average for Real subjects was 6 degrees, while Crude and Buck levels were observed with averages of 11 degrees and 9 degrees respectively. Figure 5: Interaction of Task and Method for average trunk flexion angle

5 Lateral Bending For tasks 4 to 8, the Real trials adopted postures with slightly more lateral bend of the trunk compared to Crude and Buck (Figure 6). Pooled across all of the tasks it was observed that on average the Crude method (0.8 degree) failed to result in any significant degree of lateral bending. However, there were no significant differences found between the Real or physical prop conditions. The marginal differences between Real and Buck conditions did not result in any significant difference. The following scatter plots (Figures 8 & 9) indicate the strong correlation between the real joint angles and those from the Crude and Buck physical prop levels. An overall R 2 value for Real-Buck conditions is 0.87 vs for the Real-Crude conditions, which indicate that the Buck physical conditions evoked a slightly stronger correlation with the Real subjects. Figure 6: Interaction of Task and Method for average trunk lateral bend angle Figure 8: Scatter plot indicating the relationship between Real and Crude level joint angles Axial Rotation There were no significant differences in the averages of trunk rotation between Real, Crude and Buck (Figure 7). However, with the exception of tasks which involved extended reaches (Tasks #2, 9, 10, 11, 12), the Crude and Buck levels did not exhibit any appreciable axial rotation deviations. Conversely, the Real subjects ranged from +20 to -20 degrees of axial trunk rotation across all of the tasks. Figure 9: Scatter plot indicating the relationship between Real and Buck level joint angles. Figure 7: Interaction of Task and Method for average trunk axial rotation angle

6 Table 2: Summary of RMS errors (Real versus Buck or Crude) and differences (Crude versus Buck) for the various joint angle variables. A composite score is calculated using the elbow, shoulder and trunk (averaged across the three axes) angles. RMS Error Joint Angle Real vs Crude Real vs Buck Crude vs Buck Elbow Weight Resultant Shoulder Trunk Flexion/Extension: Angle Trunk Lateral Bend: Angle Trunk Axial Rotation: Angle Average of Elbow, Shoulder and Trunk Figure 10: Comparison of all three Prop Levels indicating that errors with Real were similar for Buck and Crude, and that the differences between Buck and Crude were relatively small across the 12 tasks. DISCUSSION Elbow angle was found to have small differences between Real and the other physical prop levels (RMS error). Real was generally found to have the most weighted elbow flexion (42 degrees) as compared to Crude and Buck conditions which both averaged degrees of elbow flexion. The elbow flexion angles were consistently flexed across all tasks performed at a close horizontal reach from a generally upright trunk posture (Tasks #1, 8, 9). There was inconsistency across the one or two-handed overhead working postures. Task #4, which involved loading a part to the underneath of the vehicle, resulted in similar weighted elbow flexion for both all conditions. To the contrary, the Buck level differed from the Crude and Real conditions for Tasks #5 and 6. For these tasks, the elbow angle was generally less flexed and there was variation as to where each arm was placed. There were also some inconsistencies across tasks which required extended horizontal reaches (Tasks #2, 3, 10). For task #2 the Real and Buck levels were characterized by more elbow flexion while the Crude level accomplished the task with an extended elbow. Yet, for the extended reaches at lower heights, the Crude and Buck levels were observed to have a higher of degree of trunk flexion and reduced lateral bending and therefore were able to complete the task with a flexed elbow. Under the same kinematic constraints, the Real subjects adopted a fully extended elbow posture to compensate for less trunk flexion. Finally, the Real subjects performed the one-handed above shoulder tasks, completed inside the vehicle, with substantially more elbow flexion than the other conditions which were almost fully extended, especially in Tasks #11 and 12. All methods were usually found to have similar amounts of forward-backward and abduction-adduction shoulder rotation for each task. When pooled across all of individual tasks the weighted resultant shoulder rotation which averaged 112 degrees of rotation for the Real subjects and degrees for both the physical prop levels. With the exception of Tasks #1, 7 and 8 where the hand positions were almost ideal, the Buck conditions evoked postures with more shoulder abduction than the Real subjects. Across all extended reach postures (Tasks # 2, 3, 9, 10, 11, 12) the Crude and Buck conditions would have more trunk flexion and shoulder rotation, while the Real subjects used more lateral trunk bend, less shoulder flexion and more elbow extension. This was also consistent for the one and twohanded tasks (Tasks #4, 5, 6). During the Crude and Buck conditions, subjects stood directly under the task hand location and subsequently incurred more shoulder abduction-adduction rotation (120 degrees), in comparison to the Real subjects (110 degrees). As indicated, there was good consistency between the Real and different physical prop levels for tasks that were performed in a relatively proximal horizontal reach (Tasks # 1, 7, 8). Most of these tasks were performed in an upright posture, with some varying degrees of forward flexion, yet relatively little difference in lateral bending and axial rotation was observed. For one or two-handed overhead reach tasks (Tasks # 4, 5, 6), all of the methods were observed with an upright posture, with little lateral bend and some extension. It is also interesting to note that Real workers chose to axially rotate the trunk an average of 11 degrees, while the Crude and Buck physical prop levels performed these tasks with trunk extension only (~0 degrees of axial rotation). The tasks which evoked the greatest amount of variability in trunk flexion were those that required extended reaches (Tasks #2, 3, 8, 9, 10, 11, 12). Real subjects performed the lowest horizontal reaches (Tasks #3, 8, 9) with a moderate degree of trunk flexion in combination with a fully extended elbow, while a greater degree of flexion with degree of weight elbow flexion was observed for the Crude and Buck levels. For

7 tasks which required an extended horizontal reach at higher vertical height (Tasks #2, 10, 11) the Real subjects appeared to position themselves such that the trunk was laterally bent, with a moderate degree of flexion or axial rotation. Conversely, the Crude and Buck levels showed progressively more trunk flexion, less lateral bend and more axial rotation towards the task hand. It can be inferred that the subjects performing tasks at the Crude and Buck physical prop level primarily use a combination of trunk flexion and axial rotation and to a lesser extent lateral bending to align themselves to the task. Compared to Real trials, the Crude and Buck physical prop levels tend to result in more trunk flexion, shoulder rotation and arm extension. It is hypothesized that the experience of the subjects and re-creation of the assembly environment for the Real condition evoked an opportunity to reduce deviations about the trunk and shoulder that is not accurately reflected in the other physical prop conditions. For the comparison of the postural differences evoked under the different physical prop levels, the task hand position was also considered (Figures 11 & 12). Quantifying the task hand position, with respect to the distance from hand to shoulder or L5/S1, provided an important reference to the force moment arm. Motivated by the high priority of low-back and shoulder injury across automotive assembly, such dependent variables provide significant predictors of shoulder and low-back postural parameters. It is evident from the plots that neither the Crude or Buck physical prop levels evoke significantly different moment arms with respect to either the shoulder or L5/S1 (Figures 11 & 12). Figure 12: Estimate moment arm from the hand force vector to L5/S1 In summary, when comparing the three conditions to each other, the RMS differences with Real were only slightly lower for the Buck prop level as compared to the Crude prop levels (Table 2; Figure 10). However, there was insignificant difference in the RMS error when the data were collapsed across all postural kinematic data, with the Crude and Buck physical prop levels having an error score 13 and 12 degrees respectively. It was found that one handed overhead Tasks appear to benefit from Buck props, especially in cases like Tasks 10, 11 and 12 where the other hand would likely be needed for support. Such props must be stable to allow subjects to fully support the body weight as necessary. There appeared to be no benefit of the Buck conditions over the Crude conditions for overhead tasks that use two hands (and don t require a support hand). For one handed tasks being done below shoulder height, there is a need for a physical support for the other hand, although more complex props are not necessary. CONCLUSION An effort was made to determine if there were any specific characteristics consistently associated with the motion capture of assembly tasks where physical prop level s differed significantly from the real assembly conditions. Postural measures have demonstrated that there are no significant systematic differences between Crude, Buck or Real physical prop conditions. ACKNOWLEDGMENTS Figure 11: Estimate moment arm from the hand force vector to the shoulder The authors of this paper wish to thank Motion Analysis Corp. for their generous donation of 8 Hawk cameras for the duration of this study.

8 ADDITIONAL SOURCES Jack software is copyright of Siemens Product Lifecycle Management Software Inc. Additional information about Jack can be found at: /tecnomatix/human_performance/index.shtml EVaRT Is trademark of Motion Analysis Corporation. Additional information can be found at: REFERENCES 1. Chaffin, D.B, Andersson, G.B.J and Martin, B.J. Occupational Biomechanics. Wiley, New York, Toronto, (2006). 2. Chiang, J., Stephens, A. and Potvin, J. Retooling Jack s Static Strength Prediction Tool. Society of Automotive Engineering, , (2006). 3. Classic JACK TM Commercially available Digital Human Modeling Software from UGS (Siemens, UGS, Germany). 4. Godin, C.A., Chiang, J., Stephens, A. & Potvin, J. Assessing the Accuracy of Ergonomic Analyses when Human Anthropometry is Scaled in a Virtual Environment. Society of Automotive Engineering, , (2006). 5. Regents of the University of Michigan (2000). University of Michigan 3DSSPP Stoffregen, T.A., Bardy, B.G., Smart, L.J., and Pagulayanm R.J. On the nature and evaluation of fidelity in virtual environments. In L.J. Hettinger and M.W. Haas (Eds.) Virtual and Adaptive Environments: Applications, Implications and Human Performance Issues, Mahwah, NJ: Erlbaum, , (2003). 7. Wilson, J. R. Virtual environment applications and applied ergonomics. Applied Ergonomics, 30, 3-9, (1999).

9 WS6: Securing Muffler and Mid-Pipe WS5: Alignment of Exhaust System WS2: Dash panel Grommet Install WS3: Locating Thin Foam Pad WS4: Loading Mid-Pipe to Muffler WS1: Hood Release Cable Install APPENDIX WS9: Electrical Connection WS8: Installing PIA Pushpins to Frame WS7: Installing PIA Pushpins to Frame WS12: Seating Rear Headline Pushpin WS10: Seating Headline Retaining Pin