### Abstract

- Top of page
- Abstract
- 1. INTRODUCTION
- 2. METHODS
- 3. RESULTS
- 4. DISCUSSION
- 5. CONCLUSIONS
- ACKNOWLEDGMENTS
- References

Video-based posture analysis employing a biomechanical model is gaining a growing popularity for ergonomic assessments. A human posture simulation method of estimating multiple body postural angles and spinal loads from a video record was developed to expedite ergonomic assessments. The method was evaluated by a repeated measures study design with three trunk flexion levels, two lift asymmetry levels, three viewing angles, and three trial repetitions as experimental factors. The study comprised two phases evaluating the accuracy of simulating self- and other people's lifting posture via a proxy of a computer-generated humanoid. The mean values of the accuracy of simulating self- and humanoid postures were 12° and 15°, respectively. The repeatability of the method for the same lifting condition was excellent (∼2°). The least simulation error was associated with side viewing angle. The estimated back compressive force and moment, calculated by a three-dimensional biomechanical model, exhibited a range of 5% underestimation. The posture simulation method enables researchers to quantify simultaneously body posture angles and spinal loading variables with accuracy and precision comparable to on-screen posture-matching methods.

### 1. INTRODUCTION

- Top of page
- Abstract
- 1. INTRODUCTION
- 2. METHODS
- 3. RESULTS
- 4. DISCUSSION
- 5. CONCLUSIONS
- ACKNOWLEDGMENTS
- References

As indicated by several review studies summarizing a significant relationship between poor working posture and the development of musculoskeletal disorders (MSDs) (Bernard, 1997; da Costa & Vieira, 2010; Ferguson & Marras, 1997; Hoogendoorn, Poppel, & van Bongers, 1999; National Research Council, 2001), body posture has been a focus of ergonomic assessments. In particular, trunk flexion and twisting/asymmetry have been demonstrated to be significant risk factors for low back disorders (LBDs) (Hoogendoorn et al., 2000; Jorgensen, Marras, Gupta, & Waters, 2003; Marras et al., 1995; Punnett et al., 1991). In most epidemiological studies, working posture is typically recorded by self-administered questionnaire or pencil/paper observational methods (Burdorf, 1992; Li & Buckle, 1999). Because of the nature of the methods, assessments of body posture have been mostly described in gross categorical terms, resulting in relatively moderate associations with LBDs (Marras, Lavender, Ferguson, Splittstoesser, & Yang, 2010). Misclassification of the gross terms as physical risk factors for LBDs raises questions about their validity and relationship with LBDs (Punnett & Wegman, 2004). Epidemiological evidence associated with LBDs may be identified when the physical risk factors are properly addressed by biomechanical factors, such as the load location and weight magnitude relative to the worker and three-dimensional movements during lifting (Burdorf, 1992; National Research Council, 2001; Sutherland, Albert, Wrigley, & Callaghan, 2008; Marras et al., 2010; Boda, Bhoyar, & Garg, 2010).

In lieu of field-friendly, direct-reading measurement methods for body posture, computerized video-based posture analysis has been favored by many researchers as a practical alternative (Bao, Howard, Spielholz, & Silverstein, 2007; Callaghan, Salewytsch, & Andrews, 2001; Keyserling, 1986; Yen & Radwin, 1995). The validity of a video-based posture analysis primarily depends on the assessed body angle (Bao, Howard, Spielholz, Silverstein, & Polissar, 2011; Burt & Punnett, 1999; Ericson, Kilbom, Wiltorin, & Winkel, 1991; Genaidy, Simmons, Guo, & Hidalgo, 1993; Lau et al., 2010; Lowe, 2004; Lu, Waters, Werren, & Piacitelli, 2011; Xu & Chang, 2011) and posture-viewing angle (Bao et al., 2011; Lu, Waters, Werren, & Piacitelli, 2009, 2011; Xu & Chang, 2011). The main advantage of this analysis method is minimal disruption to workers’ job performance during field surveillance with a permanent record for future analyses at a very low cost (Bao et al., 2011; Genaidy et al., 1993; Li & Buckle, 1999). Recorded posture or movement data can be further used for biomechanical modeling (i.e., inverse dynamics) to obtain joint loading variables (Chaffin, 1969; de Looze, Kingma, Bussmann, & Toussant, 1992; Kromodihardjo & Mital, 1986; Kingma, de Looze, Toussaint, Klijinsma, & Bruijnen, 1996) and cumulative spinal loads (Callaghan et al., 2001; Jager, Jordan, Luttmann, & Laurig, 2000; Kumar, 1990; Lu et al., 2011; Norman et al., 1998; Sutherland et al., 2008). However, quantifications of cumulative spinal loading variables involve laborious manual mannequin/stick figure manipulation or manual screen digitization of body joints to match the posture on the computer screen, which is time consuming and prone to errors (Callaghan et al., 2001; Liu, Zhang, & Chaffin, 1997; Lu et al., 2011). It could take 11 min to manipulate the posture on the computer screen to match a working posture in a photograph (Chaffin, 1997).

To expedite this biomechanical modeling process, we developed a human posture simulation method that could simultaneously estimate multiple body posture angles from field-recorded video (Waters, Lu, Werren, & Piacitelli, 2011). The biomechanical model used in this method was the University of Michigan three-dimensional static strength prediction program (3DSSPP) (Chaffin, Andersson, & Martin, 1999; Garg & Chaffin, 1975). Using anthropometry, hand load, and posture data, this biomechanical model has the capability of predicting spinal compressive force acting at the L4/L5 intervertebral disc for a static working posture in the three-dimensional directions (Chaffin, 1969; Chaffin & Baker, 1970; Garg & Chaffin, 1975; Chaffin & Erig, 1991). The model has been widely used in many studies as design criteria for manual materials handling jobs or a risk assessment tool for LBDs (Chaffin, 1997; Garg & Kapellusch, 2009; Lavender, Oleske, Nicholson, Andersson, & Hahn, 1999; Marras, Fine, Ferguson, & Waters, 1999; Waters, Putz-Anderson, & Baron, 1998).

This article describes the development of this human posture simulation method with a goal to answer the following research questions:

- Can human subjects simulate or mimic their own and others’ posture accurately and precisely?
- How efficient is the human posture simulation method?
- How accurate is the human posture simulation method in estimating the back compressive force and moment in the lumbosacral region?

### 4. DISCUSSION

- Top of page
- Abstract
- 1. INTRODUCTION
- 2. METHODS
- 3. RESULTS
- 4. DISCUSSION
- 5. CONCLUSIONS
- ACKNOWLEDGMENTS
- References

To validate the posture simulation method, we started with the best-case scenario by simulating the subjects’ self-postures followed by simulating computer-generated mannequin postures as a proxy for other people's postures. Finally, the posture simulation errors and some nonposture variables (i.e., hand load and 50th percentile male anthropometrics) were taken into account and evaluated together during the estimation of the back compressive force and moment.

Findings from the two phases of this study suggest that humans have a great potential for simulating their own and other people's posture with reasonable accuracy and precision. The subjects demonstrated an average 12° error for simulating their own posture and an average 15° error for simulating mannequin postures for the same variety of postures in different viewing angles. The small ∼2° mean value of the precision measure from both experiments indicates an excellent repeatability of the posture simulation method for simulating the same working condition, which agrees with the majority of the observational methods for posture specification (Takala et al., 2010). In a study using a manual observational method to estimate the same body angles for 3DSSPP, the average posture specification error (calculated with the same method used in the present study) for a similar setting (i.e., one photograph) was about 9° (Liu et al., 1997). It is difficult to have a direct comparison in the posture specification errors between the current study and Liu's study, which involved 4 different working postures of 4 different university students. We conducted a mini-study to allow a direct comparison between the observation and posture simulation methods.

In the mini-study (Lu et al., 2009), the simulators’ self-photographed postures were presented to 5 certified professional ergonomists for specifying the trunk flexion angle on a computer screen. Results showed that the average errors in simulating and manually specifying the trunk flexion angle were about 6° and 13°, respectively. This previous study also showed an improved inter-rater correlation coefficient (ICC = 0.82 for simulation vs. 0.65 for expert rating) for estimating the trunk flexion angle (Lu et al., 2009). Figure 3 shows posture specification errors between the simulation and observational methods. As seen in Figure 3, the accuracy levels of simulating mannequin postures and expert rating are comparable, while the accuracy of simulating self-postures increases by an average 5°. The finding suggests that simulating self-posture or a worker's posture in similar type may result in a significant decrease in posture specification errors by approximately 50%. A recent study has shown an average 9° error in estimating workers’ trunk flexion angle on the sagittal plane (i.e., side view) for a variety of lifting tasks using the manual posture specification on the computer screen (Xu & Chang, 2011). It is worth noting that the 9° error falls in between the 6° and 12° errors in simulating self- and mannequin postures in our studies.

Results from three experiments in Figure 3 (simulating self-posture, mannequin posture, and rating by professional ergonomists) suggest the same trend that estimating an increased trunk flexion angle was associated with an increased estimation error. The finding is in line with several studies where increased errors in estimating shoulder and wrist posture were associated with an increased targeted angle (Genaidy et al., 1993; Lowe, 2004). Tables 2 and 3 indicate that for the posture simulation method, the error trend applied to both trunk lateral bending and flexion. The reason that the trunk axial rotation error was not associated with the increased trunk flexion angle might be attributed to the relatively smaller trunk axial rotation angles for the trunk flexion Group A (increased trunk flexion group). The trunk axial rotation angles for trunk flexion Group A were 3°–27°, while the angles for the other two groups ranged from 0° to 45°. Caution should be exercised when using the subjective rating or the human posture simulation method to estimate an increased trunk flexion angle for postural risk assessments.

Results from the Phase 2 experiment show that the mean of the simulation errors in trunk flexion, lateral bending, and axial rotation angles ranged from 4° to 18° for the three groups of trunk flexion and 8° to 16° for the three viewing angles. The results about estimating the three important postural risk factors for LBDs are not entirely satisfactory but perhaps tolerable for epidemiological research. If the simulator's body type is a match or similar in terms of height, weight, and gender to the person in the video for assessment, the posture simulation method may offer a promising approach for a fairly accurate estimation of these trunk posture variables, as indicated by an improved average 6° simulation error for a variety of trunk postures in different views in the Phase 1 experiment.

Among the three viewing angles, the side-viewing angle had the least average errors for estimating both self's and mannequin's body posture angles. According to the correlation analysis (Table 4), the estimated back compressive force and moment data for the side viewing angle was also found to have the strongest correlation (*r* > 0.9) with both self's and mannequin's. This increased correlation for the side-viewing angle seemingly was the results of the decreased average posture simulation error with this viewing angle. Therefore, to improve the accuracy of estimating the back compressive force and moment using the human posture simulation method, the side-viewing angle is recommended.

With the limited sample size used in the Phase 2 experiment, it is difficult to conclude an under- or overestimation of the back compressive force and moment using the human posture simulation method. According to the regression coefficients of the regression analysis for the 5 subjects, only 1 subject had a regression coefficient greater than 1. Most subjects had a coefficient ranging from 0.95 to 1. Therefore, the human posture simulation method appears to have a trend of underestimating both back compressive force and moment within a range of 5%.

As the level of hand load increased, the correlation between simulation and mannequin data decreased, indicating a compromised accuracy of estimating back compressive force and moment at an increased level of hand load. The decreased correlation is attributed to the multiplication effect of the horizontal moment arm from the load to the L4/L5 and hand load for calculating both force and moment data (Chaffin et al., 1999). This limitation may not be critical when one is to assess the biomechanical model-driven back compressive force and moment for small to median hand loads.

As can be seen in Figure 2, the main advantage of the posture simulation method is its efficiency in completing posture specification of the 15 body angles. Previously, we reported an average time period of 20 s to complete one posture simulation trial (Lu et al., 2011). The time was based on the preparation time and actual posture posing time (standardized 3 s) without considering the time for presenting the posture on the screen and time for saving data to the computer hard drive by the computer operator. Taking these extra factors into account, the average time (∼1.5 min) for each simulation trial is about 5 min less than manual posture-matching on computer screen in the current study. Because of the limited data (one subject's 54 trials) we collected for assessing the efficiency of the on-screen posture-matching method, the comparison does not seem generalizable. However, the time required for completing one human posture simulation trial is on average 3 min less than the reported data for similar on-screen posture specification trials reported in previous studies (Liu et al., 1997; Xu & Chang, 2011). For a large-scale epidemiological study involving measuring the whole-body postures for many manual materials handling tasks, the human posture simulation method may offer an efficient, precise, and reasonably accurate way.

Our human posture simulation method has two advantages over on-screen manual posture-matching methods. First, the human posture simulation method has the capability of specifying many body angles of interest at once, as compared to judging one angle at a time on the computer screen using the manual posture-matching methods. This advantage leads to a significant reduction in labor and analysis time required by the on-screen methods. Second, human posture simulation has the capability of linking many body angles realistically, as compared to specifying one body angle at a time without considering anatomical limitations between the body angle being specified and other body angles using the on-screen methods. For example, excess trunk axial rotation typically comes with pelvic rotation in the same direction, resulting in a reduction in the trunk twisting or axial rotation angle with respect to the pelvis (Anderson, Chaffin, & Herrin, 1986). This reduction is difficult to take into account during manual posture matching on the computer screen and typically is not addressed by researchers. In 3DSSPP, a body linkage constraint algorithm is implemented to limit unrealistic manual posture specification; however, this constraint algorithm is based on range of motion data and is relatively unrealistic for most lifting situations, compared to human posture simulation that employs a real human body composed of natural constraints from muscles, bones, ligaments, and soft tissues.

The study methodology has some limitations that warrant considerations. First, the small sample sizes for both experiments provide limited generality. Although statistically significant differences were found for the average posture simulation error in both experiments, the nonsignificant associations between test conditions (viewing angle, lift asymmetry, and simulated trunk flexion angle) and some trunk posture simulation errors may raise a question about the appropriate sample size for assessing the simulation errors in these trunk posture variables. Post hoc power sample size calculations were performed to estimate the required sample sizes for the nonsignificant trunk posture variables for both experiments. To detect a significant difference (*p* < .05 with power = 0.9) in the accuracy of simulating the trunk lateral bending, axial rotation and flexion variables in the test conditions in the Phase 1 experiment, a minimal sample size of 19, 26, and 41 was required, respectively. For the same trunk variables with the same statistical power in the Phase 2 experiment, a minimal sample size of 13, 21, and 14 was needed, respectively. Research with a larger sample size is recommended to provide further data to quantify the human posture simulation errors. Second, theoretically, the simulation error for each posture variable may be influenced by relevant posture variables. For example, the simulation error for trunk axial rotation may be affected by the simulation error for trunk flexion because of the anatomical link. However, the interactions between the simulation errors for the 15 body posture variables were not assessed in both experiments. Because of the limited sample size, an assessment of the complicated interaction terms may not justify meaningful results. A preliminary analysis of linear correlation between the trunk posture simulation errors revealed that the correlation coefficients varied from 0.2 to 0.3 for both experiments. The poor correlations imply the complicated hypothesized interactions that may warrant further investigations. Third, the computer-generated mannequin posture was not realistic; however, to investigate the effects of simulation errors on the accuracy of estimating the back compressive force and moment by the posture simulation method, it was perhaps the most accurate method for comparisons. The resolution of the mannequin for rendering a realistic human figure should have sufficient details for posture specification, as presented in Table 1. In addition, the use of the mannequin as a realistic person completely eliminated the probabilities of measurement errors using a person's posture data recorded by the motion capture system. Using a real person's data for comparisons may also raise a question about the unknown effects of personal body characteristics, such as body type, size, and height. Because of the number of trials to be completed during the test session, these personal factors were difficult to investigate simultaneously with the studied factors (viewing angle, trunk flexion level, and lift asymmetry). Nevertheless, future research on the additional factors is recommended to further validate the applications of the method.

While direct-reading measurement tools for physical risk quantifications are not well developed for field use, the human posture simulation method offers a novel approach to quantifying postural stress and biomechanical measures in the workplace. Since the method is based on the premise that biomechanical measures are associated with the development of LBDs, the value of the method in correlating the biomechanical variables with health outcomes has yet to be evaluated.