ERGONOMICS IN DIGITAL ENVIRONMENTS:DIGITAL HUMAN FIGURES

DIGITAL HUMAN FIGURES

Kinematic Representation

Human models are varied in both their complexity and construction. Any mathematical representation of human structure, physiology, or behavior can be considered to be a human model. For example, complex models of human musculoskeletal dynamics are commonly used to study motor control issues (Winters and Woo 1990). These models are typically quite detailed to allow the dynamic effects of individual muscle activation and contraction, and hypothesized neural control strategies, to be investigated. Moreover, this detail is typically focused on one part of the body, for example the lower extremity for gait analysis, or the upper limbs for investigation of movement control. In con- trast, simple, sometimes incomplete, human forms are used in the investigation of cognitive models, wherein the human form acts as an agent to effect changes in its world. The pedagogical agent developed at the University of Southern California Information Sciences Institute (Johnson et al. 2000) is an example. These models focus on the cognitive rather than motor processes and simulate the interactions among multiple humans.

For physical ergonomics investigations in digital environments, the human models need to mirror our structure, shape, and size in sufficient detail to allow the figures to assume realistically the observed postures of actual individuals performing similar tasks. Such models typically consist of an underlying kinematic linkage system that closely parallels our own skeletal structure and an attached geometric shell that duplicates our surface shape.

Today’s human models have kinematic linkages that include from 30 to 148 degrees of freedom, depending on the detail provided in the hands, feet, shoulder, and spine. The joints are constructed to move like our own joints, with the appropriate number of degrees of freedom, and typically also have physiological limits on the range of motion. In more detailed models, the shoulder and spine are modeled to behave naturally, with the links moving in concert as the particular segment is manipulated. For example, the shoulder complex consisting of the sternoclavicular, acromioclavicular, and glenohumeral joints is modeled to move in a realistic pattern when the upper arm is adjusted, moving the elevation and fore–aft position of the shoulder as the arm moves through its range of motion.

Anthropometry

The internal skeletal structure and surface topography of a digital human figure influence both the qualitative and quantitative use of the figures. As an engineering tool, the accuracy of the internal link structure affects the dimensional measures made between the environment and the human figure, such as head clearance and reach. The ability of the figure to take on physiologic surface topography directly adds to the perception of reality when one is viewing a simulation. While both of these aspects are important, to date more effort has been concentrated on the accurate scaling of the link lengths in commercial human modeling. This bias is in part motivated by the large amount of tra- ditional 1D anthropometric data available (e.g., stature, sitting height, shoulder breadth), in contrast to the largely nonexistent 3D surface contour data available. Secondly, a driving factor of human modeling in visualization environments has been to produce a system that works in near real time (Badler et al. 1993). The complexity of the figure surface description presents a burden on the real- time performance, so a balance is sought in which there is sufficient surface detail for visual reality without undue computational load. As computer hardware technology improves, the ability to add to this surface complexity is afforded.

Anthropometric Databases

Of the many anthropometric databases available, one of the most widely used is the U.S. Army 1988 Anthropometric Survey (ANSUR) (Gordon et al. 1988). The ANSUR study was performed by the U.S. military to provide a representative anthropometric database of the U.S. military personnel. This database has a demographic representation that matches the U.S. army, which is known to differ from the gender, racial, age, and conditioning distributions of the population as a whole. Nevertheless, the statistical measures of the ANSUR data have been estimated to be within 3% of the general U.S. civilian population (Roebuck 1995). This study contains 132 standard anthropometric measurements from approximately 9000 military personnel, of which a sample of 1774 men and 2208 females were selected to represent accurately the military population demographics. Documents that contain the individual subject data as well as the summary statistics are publicly available, so publishers of human modeling software can independently develop statistical models for figure scaling and boundary manikin generation.

Another anthropometric database available is the National Health and Nutrition Examination Sur- vey III (NHANES 1994), which contains the dimensions of 33,994 persons ages 2 months and older, of which 17,752 are age 18 and older. While the 21 measures of this database do not provide enough information to define adequately the dimensions of most contemporary human models, the database currently represents the most comprehensive and representative database for the U.S. population. These publicly available data contain weighting information based on the most recent U.S. census (1988–1994). The census weighting data allow U.S. representative population statistics to be com- puted for any population selections based on gender, race, ethnicity, and age. While both the ANSUR and NHANES data describe single dimension measures taken between anthropometric landmarks, a new anthropometric survey has been initiated to provide a database of population 3D body shapes. The CAESAR project (Civilian American and European Surface Anthropometric Resource) will scan approximately 6000 individuals in the United States and Europe. These data are in the form of both traditional anthropometric measures and new 3D data from whole body laser scanners, that provide a highly detailed data cloud describing the shape of the subject surface contour (Figure 1).

Both children- and nationality-specific anthropometric databases are also available, although these databases have not been adopted to the same degree as those previously mentioned due to their limited international availability and data restrictions (Table 1).

Accommodation Methods

One of the advantages digital ergonomics can bring to the development process is the ability to investigate accommodation issues early in the design process. In the past, physical mockups were created and evaluated using a large subject population to arrive at accommodation metrics. This approach is both expensive and time consuming and does not lend itself to rapid evaluation of design alternatives. In the digital space, a population of figures can be used to investigate many of the same

issues of clearance, visibility, reach, and comfort. Defining the sizes of the manikins to be used in the process is one of the first steps of these analyses.

To perform an accommodation study, the user defines a percentage of the population that he or she wishes to accommodate in their design or workplace and then scales representative figures using data from applicable anthropometric databases to represent the extreme dimensions of this accom- modation range. As few as one measure, often stature, may be judged to be important to the design and used in the creation of the representative figures. For other applications, such as cockpit design, multiple measures, such as head clearance, eye height, shoulder breadth, leg length, and reach length, may all affect the design simultaneously. Several methods are in use to investigate the accommodation status of a design and are described below.

Monte Carlo Simulation The Monte Carlo approach randomly samples subjects from an anthropometric database to create a representative sample of the population and processes these figures through the design. Recording the success or failure of the design to accommodate each

subject data

individual of the sample allows an indication of the percentage accommodation to be derived. This method is computationally fairly expensive because it requires that a large number of figures be generated and tested for a meaningful analysis. Also, because the distribution of sizes follows a bell- shaped distribution, many more people are close to the average than to the extremes, which results in many figures of fairly similar dimensions needlessly tested. This can make this approach somewhat inefficient.

Boundary Manikins In contrast, the boundary manikin approach can be used for mul- tiple dimensional analysis (Zehner et al. 1993; Bittner et al. 1986). The statistics of principal com- ponents (factor analysis) can be used to identify bounding multidimensional ellipsoids that contain a portion of the population. For example a 95% hyperellipsoid can be found that defines the dimen- sional ranges for all variables of interest within which 95% of the population can be expected. This approach can reduce the number of manikins that need to be tested in a design to a practical number, which is beneficial in cases where the computer rendering speed or the effort to manually posture the figure is significant.

Whole-Population Analysis Both the Monte Carlo and boundary manikin approaches attempt to reduce the number of subjects that are run through the analysis while still providing statistical validity to the results. However, as computer technology improves and as models to posture the manikins realistically in the environment become available, it becomes not unreasonable to run several thousand figures through a design automatically. Because this approach does not use statistical models of the data but instead uses the measured data directly, the unexplained variability that is not captured by the statistical data modeling is avoided. This approach still requires substantial run time and is currently not a popular approach to accommodation analysis.

Human Figure Posturing

As mentioned briefly in the previous anthropometric discussion, figure posturing is a critical com- ponent, along with anthropometry, in accommodation studies. While automatic posturing methods based on empirical studies are becoming available and will be discussed in later sections, there are more fundamental lower-level tools that have been developed for general manipulation of figures in the virtual environments. Because contemporary figures may have over 100 DOF, adjustment of each joint angle individually is unworkably tedious.

Coupled Joints

Fortunately, the human skeleton, while infinitely adjustable, is held together by muscles and connec- tive tissues that constrain the movement of certain joints to work as motion complexes. Examples include the shoulder, spine, and hands. While there are obvious freakish exceptions, most people cannot voluntarily dislocate their shoulders or move the carpal bones in the digits independently. Modelers take advantage of these relationships to build movement rules for these joints such that many degrees of freedom can be adjusted easily with a few angles that are defined in common human factors parlance. For example, the EAI Jack human model has 54 DOF in the spine below the neck, which can be manipulated using three standard angle definitions: flexion, extension, and axial rotation. Similarly, as described earlier, the shoulder comprises a clavicle link that moves in concert with the arm as it is manipulated, mimicking the natural kinematics of this joint. Such coupled joint complexes greatly simplify the posturing of high-degree of freedom human figures.

Inverse Kinematics

Even with the substantial reduction in degrees of freedom that coupled joints bring, there are still far too many degrees of freedom remaining in a contemporary figure for rapid posturing in production use. To address this, human modelers have drawn from the robotics field the concept of inverse kinematics (IK) or specifying joint kinematics based on a desired end-effector position. Inverse kin- ematics operates on a linked chain of segments, for example the torso, shoulder, arm, forearm, and wrist, and, given the location of the distal segment (i.e., hand), solves all of the joint postures along this chain based on some optimization criteria. For human models, these criteria include that the joints do not separate and that the joint angles remain within their physiological range of motion. Using inverse kinematics, the practitioner is able to grab the virtual figure’s hand in the 3D visual- ization environment and manipulate its position in real time while the rest of the figure modifies its posture (i.e., torso, shoulder, arm) to satisfy the requested hand position. While the IK methods can be made to respect the physiologic range of motion limitations inherent to the joints, they tend not to have the sophistication always to select the most likely or physiologically reasonable postures. This is especially problematic when the number of joints in the joint linkage is large. If the number of degrees of freedom is too great, there is unlikely to be just one unique posture that satisfies the specified end-effector position. For specific cases, this is being addressed with empirical-based pos- turing models, which are discussed in greater detail below. However, even with the caveat that IK sometimes selects inappropriate postural solutions, it is currently the most popular and rapid method of general postural manipulation in 3D environments.

Motion / Animation

While static posturing is often sufficient to analyze many ergonomic issues, such as reach, vision, clearance, and joint loading, often figure motion in the form of an animation is important. Examples include simulated training material, managerial presentations, and analyses that depend on observa- tions of a person performing an entire task cycle, for example when assessing the efficiency a work- place layout. Realistically controlling figure motion is without question one of the most challenging aspects of human modeling. Humans are capable of an almost infinite number of different movements to accomplish the same task. Indeed, people may use several postural approaches during a single task, for example to get better leverage on a tool or gain a different vantage point for a complex assembly. This incredible postural flexibility makes it very difficult for human modeling software to predict which motions a worker will use to perform a task. Most current animation systems circum- vent this dilemma by requiring the user to specify the key postures of the figure during the task. The software then transitions between these postures, driving the joint angles to change over time such that motions conform to correct times. A variety of mechanisms are used to perform the posture transitions, from predefined motion rules to inverse kinematics. Depending on the system, the level of control given to the user to define and edit the postures also varies, with some products making more assumptions than others. While built-in rules offer utility to the novice user, the inflexibility imposed by the system automatically selecting task postures can be restrictive and a source of frus- tration to the advanced user. In addition, the level of fidelity required in the motion varies greatly depending on the application. For applications such as the validation of a factory layout or animation of a procedure for training or communication purposes, a human motion simulation that simply looks reasonable may be sufficient. However, if sophisticated biomechanical analyses are to be run on the simulated motion, it may be necessary to generate motions that are not only visually reasonable but also obey physiologic rules. These include accurate joint velocities and accelerations, realistic posi- tioning of the center of mass relative to the feet, and accurate specification of externally applied forces.