IEEE TRANSACTIONS ON SYSTEMS, MAN, AND CYBERNETICS—PART A: SYSTEMS AND HUMANS, VOL. 35, NO. 4, JULY 2005 505 The (In)Accuracy of Novice Rover Operators’ Perception of Obstacle Height From Monoscopic Images Arjun Kumar Kanduri, Geb Thomas, Member, IEEE, Nathalie Cabrol, Edmond Grin, and Robert C. Anderson Abstract—Researchers have previously described a mobile robot, or rover, operator’s difficulty in accurately perceiving the rover’s tilt and roll, which can lead to rollover accidents. Safe mobile robot navigation and effective mission planning also require an operator to accurately interpret and understand the geometry and scale of features in the rover’s environment. This work presents an experiment that measures an observer’s ability to estimate height of distant (5–15 m) obstacles given an accurate local model (e.g., within 0–5 m of the rover), a panoramic image, and a physical mock-up of the local terrain. The experimental conditions were intended to represent a best-case scenario for a stopped rover equipped with short base-line stereoscopic cameras. The participants’ task was to extrapolate the well-modeled local geometry to monoscopic images of the more distant terrain. The experiment compared two estimation techniques. With the first technique, each observer physically indicated his or her direct estimates of the obstacle distance and height. With the second estimation technique, which we call horizon analysis, the observer indicated the position of the top and bottom of each rock on an image and the height was calculated by measuring the visual angle between the theoretical horizon and the points indicated by the observer. The direct estimation technique overestimated the height of the rocks by an average of 190%; the horizon analysis technique overestimated by 80%. The results suggest that even when provided with a rich set of supplementary and context information, rover operators have significant difficulty in vertically perceiving the scale of distant terrain. The results also suggest that horizon analysis is a more accurate method for determining the height of distant rover navigation obstacles, when the local terrain is nearly level. Index Terms—Computer interface human factors, mobile robot motion planning, mobile robots, spatial reasoning. I. INTRODUCTION M OBILE robots, or rovers, are an important tool for studying and operating in hazardous or remote environments. They are being used for police and military applications Manuscript received August 1, 2004; revised March 14, 2005. This work was supported by the NASA program Applied Information Systems Research under Grant NAG5-11981 and Grant NNG05GA51G. This paper was recommended by the Guest Editors. A. K. Kanduri and G. Thomas are with the Department of Mechanical and Industrial Engineering, University of Iowa, Iowa City, IA 52242 USA (e-mail: [email protected]). N. Cabrol and E. Grin are with the NASA Ames Research Facility, Space Science Division, Moffett Field, CA 94035-1000 USA (e-mail: [email protected]; [email protected]). R. C. Anderson is with the Jet Propulsion Laboratory, California Institute of Technology, Pasadena, CA 91109 USA (e-mail: Robert.C.Anderson@ jpl.nasa.gov). Digital Object Identifier 10.1109/TSMCA.2005.850601 [1], search-and-rescue missions [2], as well as scientific explorations near volcanoes [3], in arctic environments [4], and on Mars [5]. Two strategies for controlling rovers are telemanipulation and supervisory control. In telemanipulated systems, operators navigate by sending motor and steering commands to the rover. In supervisory-controlled systems, operators navigate by interacting with the rover’s semiautonomous system, often by suggesting intermediate goal locations or navigation tasks. In either case, the operator primarily relies on imagery collected by the rover’s cameras to perceive the remote environment and make decisions about where the rover should go and how to navigate there safely. The imagery may be provided by a live video feed on a television monitor or as a series or collection of still images in a specially designed operator interface [6]. Driving the rover remotely, however, is a difficult perceptual task. The difficulty of accurately perceiving rover attitude (pitch and roll) is described in [7] and [8]. However, few researchers have investigated the difficulty of perceiving other features of the environment from information provided by a rover. Estimating the height of distant obstacles is one such feature that is both of immediate practical relevance for planning and navigation and is relatively unambiguous and is easily measured. Misperceiving the height of obstacles is important because it can lead to inefficient and frustrating rover operations. While misperceiving rover attitude leads to rover rollover accidents, misperceiving obstacle height leads to a more subtle operation problem. If the operator underestimates the size of an obstacle, he or she may drive the rover along a particular course only to find the path blocked when the obstacle turns out to be unexpectedly large. If the rover operator overestimates the size of an obstacle, he or she may avoid a particular path, even though that path may be more efficient or productive. A more concerning aspect of the obstacle height problem, however, is that a misperception of obstacle height suggests a more general lack of veridical perception of the scale of the remote environment, for how can an operator misperceive one aspect of the environment, yet maintain a consistent, accurate representation of the other particulars? Many different factors may have a profound effect on the operator’s perception of the remote environment. Differences in camera resolution, field of view, camera height, image update rate, and rover velocity are all likely to affect the operator’s ability to accurately perceive the remote environment. Many other relevant factors are related to the environment and the operator, such as the presence of familiar objects, regular texture patterns, lighting conditions, operation training, and 1083-4427/$20.00 © 2005 IEEE 506 IEEE TRANSACTIONS ON SYSTEMS, MAN, AND CYBERNETICS—PART A: SYSTEMS AND HUMANS, VOL. 35, NO. 4, JULY 2005 operator experience. The exhaustive experimental analysis of each combination of these factors is beyond the scope of this work. Instead, we have designed the experiments reported here to demonstrate the need for greater study and understanding of this particular feature of rover operations. To that end, we have emphasized relatively simple techniques that favor experimental repeatability, focused on a well-known rover paradigm (Mars exploration) using a widely available data set, and relied on naïve research participants rather than trained rover operators. Although well-trained operators may have better performance than the participants reported here, the point of this work is to demonstrate that rover interface designers and mission planners can not assume that operators accurately perceive the remote environment, even when they are provided with a relatively rich set of information. Their ability to interpolate information to construct an accurate gestalt model may be severely limited. Furthermore, rather than simply report the estimates of the participants, we seek to explain how the observer made (or did not make) these estimates. Finally, we propose an estimation scheme that, although not without limitation, appears to be more effective than the default approach. In this manner, we hope to report a quantitative performance benchmark by which other rover systems may be compared and to begin to provide a theoretical framework within which this particular perceptual challenge may be addressed and eventually resolved. II. BACKGROUND People are not good at judging the heights of unfamiliar objects. When judging the height of unfamiliar targets in a broad, open field, people overestimate the height of the targets by as much as 225% [9]. Height estimate variability depends on the distance between the observer and the target and as well as the availability of pictorial depth cues, such as the presence of familiar objects or a clear horizon. When comparing the size of 5–15-cm targets placed on a nearby table with similar targets placed on a lawn at distances of 5, 24, and 98 m, research participants overestimated the size of the more distant targets by 25% [10]. In another experiment, participants verbally estimated the height of 30–350-cm targets at distances of 25–400 m. When the visual angle subtended by the objects was large, approximately 25 visual degrees, the participants underestimated target size by as much as 10%, but when the visual angle was small, they overestimated the target size by as much as 225% [9]. When participants were asked to compare the height of a 107-cm target placed at 30 m with similar targets placed at distances of 30–1200 m on a long runway, the participants overestimated the height of the stimulus by between 6.41% and 34.04%, depending on the target distance [11]. The presence of familiar objects and conspicuous depth cues in the environment can greatly enhance the accuracy of height judgments [12], possibly because a different processing mechanism is used when familiar objects are present [13]. Because the experiments described here use stimuli from the surface of Mars, the participants are not expected to be familiar with the size of any of the features, beyond those presented in the experiment’s physical mock-up. However, if a rover is operating in a familiar domain, which might be the case with an urban search-and-rescue rover, for example, more accurate size estimates would be expected when familiar objects are available. There are at least two competing theories about how people perceive the size of objects. The first, called the Size Distance Invariance Hypothesis (SDIH), assumes that an observer estimates height as the product of distance to the object and the vertical angle subtended by the object [14]. The alternative theory assumes that people interpret the texture gradient and other invariant visual characteristics to infer the three-dimensional (3-D) structure of the environment. The observer then estimates size to be consistent with the larger 3-D structure [15]. One invariant visual characteristic useful in the estimation of object height is the angle between the target object and the horizon. A. Size Distance Invariance Hypothesis The SDIH [16] states that an observer estimates the height of an object as the product of the visual angle represented by the retinal projection of the object , and the distance between the observer and the object (1) This hypothesis is supported by an experiment [17] that showed that height estimates covary with estimated distance and estimated visual angle. In particular, the experiment showed that (1) predicts height estimation if is taken as the estimated distance rather than the actual distance. In [16] an experiment tested the invariance hypothesis independent of any distance estimate. Participants compared visual stimuli to a set of reference objects, where both references and stimuli were displayed without distance cues. The authors argue that in the absence of distance cues, participants will perceive the stimuli and references to be at the same distance. The participants estimated the objects to be the same size when they subtended the same visual angle, consistent with the assumptions of the SDIH. Holway and Boring [17] measured stimuli sizes with four different sets of distance cues. When distance cues were completely eliminated, size estimates varied with the visual angle; when distance cues (such as binocular cues) were available, size estimates varied with stimulus height. This indicates that perception of the size of unfamiliar objects depends on both the visual angle and the perceived distance, again consistent with SDIH. Many other experiments validating the SDIH [18]–[20] have been conducted with physical stimuli, but relatively few test the SDIH with image stimuli. This may be a result of the theoretical complexity introduced by presenting a photograph rather than a physical stimulus. The SDIH research suggests that height estimation depends on distance estimation; it is not clear, however, how accurately people can estimate distance with an image B. Horizon Analysis Gibson [15] argued that people observe spatial relations through invariant structures in an optic array. Gibson favored the invariant structures present in texture gradients. For example, a monoscopic image of a homogeneously textured KANDURI et al.: (IN)ACCURACY OF NOVICE ROVER OPERATORS’ PERCEPTION OF OBSTACLE HEIGHT Fig. 1. Angles used in Sedgewick’s horizon analysis, where the observer is represented by a camera mounted on a mast. surface provides a clear cue to a surface’s geometry (but not its scale). Sedgwick [21] explored the utility of the horizon in determining object height, where horizon is defined as an infinitely distant great circle of the optic array located at the height of the observer and perpendicular to the direction of the plane upon which the observer rests. Sedgwick proved that when an object rests on a perfectly flat plane, or ground, its height may be determined from the angles between the horizon and the top and bottom of the object. This height estimate does not directly depend on the distance between the object and the observer. Sedgwick’s theory, which we refer to as horizon analysis, calculates the height of an object resting on the ground as a function of , the height of the camera focal point or observer’s eye, , the angle between the horizon and top of the object, and , the angle between the horizon and bottom of the object (see Fig. 1) (2) The principle advantage of this approach is that these parameters do not require a depth estimate. The principle limitation is that the model assumes that the ground is flat. Sedgwick [21] applied horizon analysis to the perception of both physical objects and images of objects. The extension to images, however, requires that the observer must correctly estimate the camera height and angles and . The conversion from vertical lengths in an image to a visual angle also depends on the image’s vertical field of view. Rogers and Costal [22] investigated the ability of artists and university students to use a horizon to estimate relative size in simple line drawings. Their stimuli consisted of a horizontal line representing the horizon and a vertical line representing a target extending upward from the ground plane. Their research participants drew a vertical line to indicate how large the target would appear if it were located at the designated reference position. The artist’s estimates were within –6% and 15% of the height predicted by horizon analysis, whereas the university students varied between –21% and 23%. This suggests that the ability to effectively use horizon information is a skill that varies among individuals. Rogers [23] sought to further understand whether observer height estimates were consistent with horizon analysis by manipulating the position of the horizon line in simple line 507 drawings. When the location of the horizon line was raised or lowered, the consistency of observer estimates ranged from approximately 30% overestimation to 20% underestimation, with most of the estimates consistent within approximately 5%. The results suggest that the accuracy of height estimates from images follows the pattern predicted by horizon analysis, particularly when the horizon appears near the center of the image. The evidence for SDIH is generally drawn from direct observations of test stimuli in natural environments, but not from images of these scenes, while the evidence for horizon analysis is generally from experiments with simple line drawings. The experimental evidence in both cases suggests that in some conditions people can significantly overestimate height. However, the evidence with the horizon analysis suggests that with very little information, observers can make surprisingly accurate height estimates. Neither theory defines the process an observer would use to estimate object height using all of the information provided by a rover, nor does either theory predict how accurate such estimates might be. C. Rover Operator Obstacle Height Estimation Because consistently overestimating obstacle heights limits an operator’s ability to plan rover actions beyond the range of locally modeled terrain, it is useful to consider strategies that would allow the analyst to make more accurate estimates of the height of distant obstacles. This paper compares two estimation strategies: direct, physical height estimation and a horizon analysis-based estimation. The first, which follows current estimation strategies, provides the analyst with a 3-D model of the local terrain and then requires the operator to extend this model to provide scale for more distant objects. This approach tends to support SDIH, because it emphasizes information related to distance. The second approach supports horizon analysis. The analyst need only indicate the position of the top and bottom of the object in an image, because the position of the camera, horizon and visual angles may all be calculated automatically. D. Objectives and Hypothesis The objective of the following experiment is to compare the accuracy of two methods of estimating the height of rocks 5–10 m from the camera in monoscopic images of a natural environment containing no familiarly sized objects. The first estimation method, direct estimation, is the operator’s direct physical indication of the height of the object made by marking a similar distance and height in a hallway. The second method, horizon estimation, is based on the operator’s indication of the top and bottom of the feature in an image. In all cases the operator is provided with a 3-D model of the local terrain (less than 5 m), a panoramic image, and a simple physical mockup of the environment including the position of the camera and several prominent features. The experimental hypothesis is that horizon estimation is more accurate than direct estimation because it avoids the inaccurate process of estimating distances from monoscopic images. 508 IEEE TRANSACTIONS ON SYSTEMS, MAN, AND CYBERNETICS—PART A: SYSTEMS AND HUMANS, VOL. 35, NO. 4, JULY 2005 distance and height of each target rock was calculated from the unmodified, 3-D model. Obstacle height is defined as the length of the vector that passes through the highest point of the rock, downward in the direction of gravity, to the base. The base is defined as the plane parallel to the viewing direction that intersects the local region of the ground on the left and right side of the rock. In each image, the position of the horizon was assumed to be horizontal and positioned above the image center by an angle equal to the camera’s tilt. C. Apparatus and Procedure Fig. 2. Screen shot of the original VRML model constructed from stereoscopic images taken by the IMP camera. A geometric model of the landing craft is located at the center of the image. The rover is above and to the left of the landing craft. III. METHOD A. Participants Eight members of the university community (five male, three female) aged between 18 and 60 years with normal or corrected-to-normal vision participated in the experiment. Each participant completed the experiment within 1.5–2 h. B. Stimuli The panoramic mosaic image, individual camera images, and the 3-D model used in this experiment are from the Mars Pathfinder data archive [24]. The experiment also employed a simple, full-scale, physical mock-up of a portion of the Pathfinder Landing Site terrain based on the objects and dimensions represented by the 3-D model. The 3-D model of the landing site was adapted from [25]. The original model indicated the scaled geometry of the terrain features from the lander location to a distance of approximately 14 m from the lander. However, for the purpose of this experiment, the model was modified by removing all features beyond the distance of 5 m. Fig. 2 presents a screen shot of the modified model as it appeared when rendered by a VRML browser plug-in. The color panoramic image, which showed all the features of the landing site terrain visible from the lander cameras, was also used in the experiment. The panoramic image has a resolution of 6230 1075 pixels, a horizontal field of view of 360 and a vertical field of view of 62 . A portion of this color mosaic is displayed in Fig. 3. The five training and nine experimental image stimuli each had a resolution of 258 246 pixels, with a vertical and horizontal field of view of 13.6 degrees. The five training stimuli were selected to indicate prominent rocks within 5 m of the lander, which were clearly identifiable in both the panoramic and 3-D model. The nine experimental stimuli (see Fig. 4) were selected to provide clearly identifiable rocks at distances between 5 and 10 m from the rover, which were also evident in the panoramic image. The target rocks provided a range of distances, heights and elevations from the ideal ground plane. The Each participant completed each portion of the experiment in the same order. First, each participant was taught how to use an interactive VRML browser window displayed on a 21-in monitor to navigate and explore the 3-D model. Then he or she was shown how objects in the full-scale, physical mock-up, which was located in the same room as the monitor, represented the physical scale of the model. The physical mock-up consisted of five boxes that represented the actual height, width, and relative position of five prominent rocks pictured in both the 3-D model and the panoramic image. A tripod positioned near the boxes represented the height and position of the Pathfinder lander camera. Each box was labeled with an image of the corresponding rock in the panorama. Next, each participant observed the panoramic image in a Photoshop window displayed on the monitor. The proctor again demonstrated the correspondence between rocks in the panoramic image and the boxes in the physical mock-up. Finally the participants observed the five training images, each of which corresponded to one of the boxes in the physical mock-up. The main experimental session consisted of showing each participant one of the nine experimental stimulus images and asking the participant to estimate the distance and height of one of the features indicated in the image. The participants were encouraged to look at the indicated stimulus image as well as the panoramic image and 3-D model. With each stimulus image, the participants were told the coordinates of the corresponding feature in the panoramic image and were permitted to study the available information until they were prepared to estimate the size of the indicated feature. To estimate the object height, the participant directed the experimenter to move forward or backward until the experimenter stood at a distance that the participant estimated to be the distance between the landing camera and the target. To help with this estimation, the origin for the estimate was a tripod set to the height of the lander camera. A tapemark on the floor indicated where the bottom of the image would theoretically intersect with the floor, given the camera field of view and camera tilt for each stimulus image. Although the tapemark represented a useful reference point given the geometric information available regarding the imaging conditions, each participant was reminded that the terrain in the image may not be level and that they should estimate the absolute distance between the camera and the rock. Once the participant was satisfied that the experimenter stood at the target distance, the participant instructed the experimenter KANDURI et al.: (IN)ACCURACY OF NOVICE ROVER OPERATORS’ PERCEPTION OF OBSTACLE HEIGHT Fig. 3. 509 Portion from the color panoramic mosaic image of the Pathfinder landing site, reproduced here in black and white. The original image is available in [24]. Fig. 4. Examples of stimuli used in the experiment. Target rocks are in dictated with a white cross. to indicate the height of the target rock with a marker on a meter stick. Once the participant was satisfied with the estimate position, the experimenter recorded the estimated distance and height. Finally, the participant indicated the precise pixel locations on the stimulus image that corresponded to the top and bottom of the rock. The experimenter recorded the pixel coordinates of these points. D. Experimental Design In each experimental block, a participant estimated the height, distance, and pixel coordinates of the top and bottom of the rock for each of the nine stimuli. Each subject participated in three successive, randomized blocks. IV. RESULTS The data from the three repetitions of each subject on each stimulus were averaged to produce 72 independent pairs of distance and height estimates from the direct estimates and 72 height estimates from the horizon analysis. Fig. 5 compares the direct and horizon height estimates for each stimulus. The average direct estimates varied from 126% to 579% and were, on average, 192% of the actual target height. The average horizon estimate for each stimulus ranged from 61% to 279% of the actual stimulus height, with an average overestimation of 79%. Height estimate error is the difference between each estimate and the actual value. Both the horizon estimate errors (0.14 m, 510 IEEE TRANSACTIONS ON SYSTEMS, MAN, AND CYBERNETICS—PART A: SYSTEMS AND HUMANS, VOL. 35, NO. 4, JULY 2005 Fig. 7. Inferred visual angle (degrees) derived from the direct estimates as a function of true visual angle (degrees). The error bars represent one standard deviation of the averaged estimates. Fig. 5. Comparison of the horizon estimation and direct perception techniques for estimating heights. The error bars represent one standard deviation of the averaged estimates across the subjects. Perfect estimates would lie on the lower dotted line. Fig. 8. Direct height overestimation as a function of true visual angle in degrees. The estimates are the average of each participant for each stimulus. The error bars represent one standard deviation of the estimates. m m Fig. 6. Estimated distance ( ) as a function of true distance ( ). The error bars represent one standard deviation. SD 0.22 m) and the direct estimate errors (0.30 m, SD 0.13 m) , , were significantly greater than zero ( , , respectively). and Fig. 6 presents the direct estimates of stimulus distance versus true distance. Ideally the slope of a regression of this data would be equal to one with an intercept of zero. However, a regression value of yields a slope of 0.2, an intercept of 7.3, and an value suggests that the true distance was not 0.03. The low a reliable predictor of estimated distance. Fig. 7 presents the estimated visual angle, which is the estimated height divided by the estimated distance, as a function of true visual angle. A regression of these parameters provides a value of 0.17. slope of 0.78, an intercept of 2.4 and an Fig. 8 presents height overestimation (directly estimated height divided by true height) as a function of the true visual angle of the stimulus. A regression of these two parameters has value of 0.53. Stimuli with a small visual angle tend to an be overestimated more than larger stimuli. V. DISCUSSION The results suggest that novice observers tend to overestimate the height of distance objects, at least with the viewing conditions similar to those used for Mars exploration. Even when provided with a physical mock-up of the height of the camera and several objects within 5 m of the rover, a 3-D model of the nearby terrain, and references that indicated the physical tilt of the camera for each stimulus, the participants overestimated the height of targets by 126% to 579%. When using horizon analysis, the overestimation was still significant, but smaller in magnitude. Consequently, the results support the assertion that horizon estimation is the more accurate and precise method for estimating obstacle height. Not only is the average error significantly smaller, but the average standard deviation is also smaller. The observations in this experiment are consistent with and extend previous results. The average direct observation overestimation of 126% is similar to the height overestimation range of 10%–235% reported in previous field studies [10]. The average to 23% horizon overestimation of 79% is larger than the range for nonartists looking at simple line drawings [22]. However, the stimuli in this experiment are different from line draw- KANDURI et al.: (IN)ACCURACY OF NOVICE ROVER OPERATORS’ PERCEPTION OF OBSTACLE HEIGHT ings in at least two important ways. First, the real images do not always contain the horizon within the viewing area, so angles and in Fig. 1 are larger than those used in the line drawings. Second, the simple line drawings assume that all the features lie on a uniform ground plane, whereas the terrain in the images naturally undulates, violated the assumption of horizon analysis and introducing error. The two most overestimated stimuli also had bases farther above the ground plane (0.15 and 0.25 m) than any of the other stimuli. Unfortunately the experiment was not designed to differentiate between these two possible sources of error. As a bellwether for the observer’s capability to interpret the geometry of the remote environment, the experimental results displayed in Fig. 4 suggest that there are important challenges. The observers’ estimates of stimulus distance are nearly independent of the actual distance. This suggests that the observers may have had difficulty in interpreting the pictorial depth cues in the image to infer much useful depth information. The pattern of results for direct estimation conflicts with the SDIH. As Fig. 5 illustrates, the participants’ indication of position and height is inconsistent with the visual angle presented in the stimulus. In all cases the participants’ recreation of the scene overestimated the visual angle presented in the image. One possible explanation for this pattern is that the participants misunderstood the camera’s vertical field of view. If the participants consistently underestimated the camera’s vertical field of view, they would consistently overestimate the visual angle of the stimulus. This misperception should lead to consistent, linear relationship between the estimated visual angle and the true visual angle. In this experiment, the regression was signifvalue of just 0.17, icantly different from zero, but had an which is not a reliable, predictive pattern. However, Fig. 5 suggests that with the exception of two outliers, there may be a reliable linear trend between true and estimated visual angle. Consequently, although the experiments generally refute SDIH as the sole estimation strategy, it may be that the observations are consistent with SDIH plus some other factor or combination of factors. The strongest trend in the experimental results was the overestimation of stimulus height as a function of the visual angle value of 0.53. As Fig. 6 inof the stimulus, which had an dicates, direct estimates of the stimulus height were within a factor of two when the stimulus presented a visual angle of just over 1 degree. This is similar to the pattern observed in [9]. If confirmed by other rover experiments, this factor-of-two relationship may provide a useful operational guideline, because it frames the accuracy of perception in a manner that is relatively easy to remember. However, this relationship is less useful to designers because it does not suggest how to improve operator’s perception of stimulus height, because the absolute visual angle presented by a target can not be manipulated. A practical application of height estimation during rover operations is to allow an operator to decide if a distant object poses a navigation threat to the rover. The Pathfinder rover was able to safely traverse obstacles less than 0.25 m [26]. Based on this criterion, seven of the nine stimuli represented obstacles that did not pose a threat to rover navigation. However, with direct estimation, all of these objects would have been perceived as 511 a threat. With horizon estimation, only one stimulus would be classified as a threat, three would be classified as not a threat, and the other three would have been classified as a threat by some analysts and not a threat by others. Consequently, although horizon estimation is still not a perfect solution for long-range rover obstacle estimation, it provides a practical improvement over current estimation techniques. Accurate estimation of obstacle height with a rover still poses a significant challenge for rover operators. Generally, current rover designs have addressed this by equipping rovers with stereoscopic camera systems. In cases where the ground is reasonably level or when a terrain model is available and the rover is aware of its position within the terrain, horizon analysis may provide an effective alternative to stereoscopic imagery. We speculate that adding graphical enhancements to provide a horizon-analysis reference may help rover operators to perceive obstacle height with reasonable accuracy, even accurate depth information is not available. Testing this speculation, however, will be the subject of future research. VI. CONCLUSION Naïve observers are not reliable judges of the height of distant objects pictured in still, monoscopic images under conditions similar to those used in Mars rover exploration. In these conditions, horizon analysis provides a more accurate and precise height estimate than direct estimation. With the conditions tested here, which included providing information to support an accurate physical understanding of the local m geometry of the camera and local terrain, observers overestimated stimulus height by an average of 192%. Direct estimation is not consistent with the SDIH or with any single source of error, such as a consistent misestimation of target distance, camera field of view, or camera height. Horizon estimation, on the other hand, is a more reliable technique, so long as the terrain is generally flat. For the conditions investigated here, horizon analysis yields a superior estimation when the bottom of the obstacle is within 10 cm of the ideal ground plane. These results suggest that designers and operators should be aware that the operator’s perception of the geometry of the rover’s environment may differ substantially from the true environment. The operator may tend to overestimate both the height and the distance of objects in the remote environment. Providing a 3-D model of the local environment does not appear to eliminate this overestimation tendency. Generally, when direct sensor measurements are not available and until better analysis techniques are developed, rover operators should avoid making navigation decisions based on the perceived height of distant objects. When operators must make such judgments, they should use horizon analysis rather than direct estimation, particularly if the terrain is generally level. ACKNOWLEDGMENT The authors would like to thank the participants of the experiment and Z. Xiang and the members of the GROK Laboratory, University of Iowa for their technical help and encouragement throughout the project. 512 IEEE TRANSACTIONS ON SYSTEMS, MAN, AND CYBERNETICS—PART A: SYSTEMS AND HUMANS, VOL. 35, NO. 4, JULY 2005 REFERENCES [1] M. H. Bruch, G. A. Gilbreath, J. W. Muelhauser, and J. Q. Lum, “Accurate waypoint navigation using nondifferential GPS,” in AUVSI Unmanned Syst., Lake Buena Vista, FL, Jul. 9–11, 2002. [2] J. Casper, M. Micire, and R. Murphy, “Issues in intelligent robots for search and rescue,” SPIE Ground Veh. Technol. II, vol. 4, pp. 41–46, 2000. [3] J. E. Bares and D. S. Wettergreen, “Dante II: Technical description, results and lessons learned,” Int. J. Robot. Res., vol. 18, no. 7, pp. 621–649, 1999. [4] D. Wettergreen, B. Shamah, P. Tompkins, and W. L. Whittaker, “Robotic planetary exploration by sun-synchronous navigation,” in Proc. 6th Int. Symp. Artif. Intell., Robot., Autom. Space (i-SAIRAS ’01), Montreal, QC, Canada, 2001. [5] G. Thomas, M. Reagan, E. A. Bettis III, N. Cabrol, and A. Rathe, “Analysis of science team activities during the 1999 Marsokhod Rover field experiment: Implications for automated planetary surface exploration,” J. Geophys. Pres., vol. 106, no. E4, pp. 7775–7784, 2001. [6] T. W. Fong, C. Thorpe, and C. Baur, “Advanced interfaces for vehicle teleoperation: Collaborative control, sensor fusion displays, and remote driving tools,” Auton. Robots, vol. 11, pp. 77–85, 2001. [7] T. Heath-Pastore, “Improved operator awareness of teleoperated land attitude,” NCCOSC Tech. Rep. 1659, 1994. [8] M. Lewis, J. Wang, S. Hughes, and X. Liu, “Experiments with attitude: Attitude displays for teleoperation,” in IEEE Int. Conf. Syst., Man Cybern., vol. 2, 2003, pp. 1345–1349. [9] R. B. Joynson, L. J. Newson, and D. S. May, “The limits of overconstancy,” Quart. J. Exper. Psychol. A, vol. 17, pp. 209–216, 1965. [10] W. M. Smith, “A methodological study of size-distance perception,” J. Psych., vol. 35, pp. 143–153, 1953. [11] A. S. Gilinsky, “The effect of attitude upon the perception of size,” Amer. J. Psych., vol. 68, pp. 173–192, 1955. [12] H. W. Leibowitz and L. O. Harvey Jr, “Effect of instructions, environment, and type of test object on matched size,” J. Exper. Psych., vol. 81, pp. 36–43, 1969. [13] R. N. Haber and C. A. Levin, “The independence of size perception and distance perception,” Perception Psychophys., vol. 63, pp. 1140–1152, 2001. [14] A. S. Gilinsky, “Perceived size and distance in visual space,” Psychol. Rev., vol. 58, pp. 460–482, 1951. [15] J. J. Gibson, The Perception of the Visual World. Boston, MA: Houghton Mifflin, 1950. [16] W. Epstein, J. Park, and A. Casey, “The current status of the size-distance hypotheses,” Psychol. Bull., vol. 58, pp. 491–514, 1961. [17] A. H. Holway and E. G. Boring, “Determinants of apparent visual size with distance variant,” Amer. J. Psych., vol. 54, pp. 21–37, 1941. [18] H. P. Zeigler and H. Leibowitz, “Apparent visual size as a function of distance for children and adults,” Amer. J. Psych., vol. 70, pp. 106–109, 1957. [19] A. H. Hastorf and K. S. Way, “Apparent size with and without distance cues,” J. Gen. Psych., vol. 47, pp. 181–188, 1952. [20] E. Chalmers and Laurence Jr, “Monocular and binocular cues in the perception of size and distance,” Amer. J. Psych., vol. 65, pp. 415–423, 1952. [21] H. A. Sedgwick, “The visible horizon: A potential source of visual information for the perception of size and distance,” Ph.D. dissertation, Cornell Univ., 1973. [22] S. Rogers and A. Costall, “On the horizon: Picture perception and Gibson’s concept of information,” Leonardo, vol. 16, pp. 180–182, 1983. [23] S. Rogers, “The horizon-ratio relation as information for relative size in pictures,” Perception Psychophys., vol. 58, pp. 142–152, 1996. [24] (2003) Planetary Data System Database, Mars Pathfinder IMP Imager. NASA. [Online]. Available: http://stardev.jpl.nasa.gov/pds/index.jsp [25] (1997) High Resolution Model. Jet-Propulsion-Laboratory. [Online]. Available: http://marsprogram.jpl.nasa.gov/MPF/vrml/pathvrml.html [26] (1997) Spacecraft: Surface Operations: Rover. Jet-Propulsion-Laboratory. [Online]. Available: http://marsrovers.jpl.nasa.gov/mission/spacecraft_rover_wheels.html Arjun Kumar Kanduri is currently pursuing the M.S. degree at the University of Iowa, Iowa City. His current research interests include actuarial science and finance. He is also exploring areas that include the modeling of uncertainty events using probabilistic actuarial concepts in engineering. Geb Thomas (M’96) received the B.S. degree in physics from the State University of New York, Stony Brook, in 1991 and the M.S. and Ph.D. degrees in industrial engineering from Pennsylvania State University, University Park, in 1995 and 1996, respectively. Since 1997, he has served as an Assistant Professor in the Department of Mechanical and Industrial Engineering, University of Iowa, Iowa City. His research interests include human factors, human computer interaction, and robotics. He has published over 50 journal articles and conference papers in this and related areas. Nathalie Cabrol received the M.S. and Ph.D. degrees in planetary science from Sorbonne University, Paris, France, in 1986 and 1991, respectively. She was a Postdoctorate Researcher with the NASA Ames Research Center, with the Association Française pour L’Avancement des Sciences (Actualités de l’Hydrologie) from 1994 to 1995, then as a National Research Council Associate from 1996 to 1998. Since 1998, she has continued her work at NASA Ames through a cooperative agreement with the SETI Institute. She has participated in many large projects involving robotics for planetary exploration. Most notably, she is a Co-investigator and Science Team Member of the NASA/JPL Surface Science Operation, Moffett Field, CA, on the 2003 Mars Exploration Rover (MER) Spirit and Opportunity Missions since 2002. She has published over 120 journal articles relating to planetary geology and robotic planetary exploration. Edmond Grin received the M.S. degree in geography and geology from the University of Neuchâtel, Neuchâtel, Switzerland, in 1983, the M.S. degree in theoretical physics from the University of Lausanne, Lausanne, in 1942, the M.S. degrees in geography and land development and quantitative geography modeling from the University of Sorbonne, Paris, France, in 1985 and 1986, respectively, and the Ph.D. degree in civil engineering from the Swiss Federal Polytechnicum Institute, Zurich, Switzerland, in 1947. After a successful career as a Hydraulics Engineer from 1948 to 1975, he turned his attention toward Mars in 1987, working first as a Research Associate in planetary sciences at the Paris-Meudon Observatory, Laboratoire de Physique du Système Solaire, France, from 1987 to 1994. He then was with the NASA Ames Research Center in 1994, first as a Research Associate (1994–1998), then under a cooperative agreement with the SETI Institute. In addition to his interests in Martian hydrology, he has participated in a large number of rover field tests, often serving as the geology team’s ground truth geologist, assessing the geologic history of the site in person for comparison with the remote science team’s assessment. He has been a Collaborator with the Mars Exploration Rover Science Team since 2002. Robert C. Anderson received the B.S. and M.S. degrees in geology from Old Dominion University, Norfolk, VA, in 1979 and 1985, respectively, and the Ph.D. degree in geology and planetary science from the University of Pittsburgh, Pittsburgh, PA, in 1995. Since receiving his Ph.D., he has been with the Jet Propulsion Laboratory, Moffett Field, CA, where he specializes in Mars exploration first with the Mars Pathfinder Mission and more recently with the Mars Exploration Rovers.
© Copyright 2026 Paperzz