experiment to investigate the production of specific values of vehicle velocity. In the second experiment, subjects were ... In the case of making left or right turns, or driving on curves, the optic flow in the display of a driving simulator is ... ____ ...
OPTIC FLOW IN DRIVING SIMULATORS Ronald R. Mourant, Beverly K. Jaeger, and Yingzi Lin Virtual Environments Laboratory 334 Snell Engineering Center Northeastern University Boston, MA 02115-5000
In the case of driving simulators, optic flow is generated as a user’s vehicle traverses a three-dimensional virtual environment. The amount of optic flow is greatest directly in front and to the sides of the vehicle, with optic flow being zero at the point of expansion. In the first experiment, the method of pairedcomparisons was used to investigate perceived optic flow for four sizes of geometric field of view (GFOV) (25, 45, 65, and 85 horizontal degrees). Results established that subjects could accurately differentiate the amount of optic flow for the four levels of GFOV. In addition, subjects perceived the velocity of oncoming vehicles to be faster when the GFOV was large. We designed a second experiment to investigate the production of specific values of vehicle velocity. In the second experiment, subjects were asked to accelerate their vehicle from a stop until they perceived it was going either 30 or 60 mph. The sizes of GFOV were set at 25, 55, or 85 horizontal degrees. Subjects overestimated velocity for the 30 mph condition. When asked to produce a velocity of 30 mph they produce a velocity of 50 mph. Their production of 60 mph was much more accurate. An interaction of the velocity asked for and GFOV, showed that accuracy increased as the GFOV increased. These experiments suggest that the perception of optic flow in driving simulators is not the same as in the realworld. This results in a conflict between what is perceived from observation of optic flow in a simulated virtual environment and what is reported by the vehicle’s speedometer.
DSC 2007 North America – Iowa City – September 2007
Introduction The concept of optic flow as a person or vehicle moves through an environment was introduced by Gibson (1979). When driving, optic flow is greatest directly in front of the vehicle and decreases to zero at the focus of expansion (where the road meets the horizon). Consider a dashed lane divider on a highway. The flow of dashes appears to be faster the closer a driver looks in front of the vehicle. When looking to the left or right of the direction of travel, optic flow increases with the angle of eccentricity. Thus, if a driver looks 90 degrees to the left or right of the direction of travel (not a good thing to do) perceived optic flow is high, particularly if objects, such as trees, buildings, or other vehicles, are in close proximity. The faster the vehicle is going, the more optic flow. Thus, the amount of optic flow may serve as a natural speedometer. Evans [1970] found that drivers estimate velocity using optical flow. Lee (1980) has stated that optic flow is useful to drivers when negotiating curves and avoiding collisions with other vehicles and static objects. The perception of optic flow in driving simulators is different from that in the real world. We consider two cases: 1) driving straight ahead and 2) driving on curves and making 90 degree turns. When driving straight ahead, the real world field-of-view (FOV) of about 190 degrees is seldom available in a driving simulator. Typically, driving simulators provide a 30 to 60 degree FOV because most driving simulators use a single computer monitor as the display. In such situations, the optic flow that appears in the real world at angles greater then that provided by the simulator, is simply not available to the driver of the simulator. When driving simulators do provide a 190 degree or larger FOV, there are usually distortions in the amount of optic flow (as compared to that in the real world) due to the display not being the inside of a cylinder. Providing large FOVs (greater then 60 degrees), typically requires multiple projectors and thus edge blending and color matching are required. Few simulators do an adequate job of this. Such factors may distract from the perception of optic flow. In the case of making left or right turns, or driving on curves, the optic flow in the display of a driving simulator is always greater then that found in the real world. This is due to the vehicle actually turning in the real world. In a driving simulator, the vehicle (or buck) remains straight ahead even though the driver has turned the steering wheel 90 degrees. Thus when making a turn in a driving simulator, a driver experiences considerable more optic flow then when making a turn in the real world. A solution to this is to have a turning simulator such as that implemented by Asano and Uchida (2005). Some driving simulators use a geometric field-of-view (GFOV), where the FOV is set in the software. This allows a large FOV in the 3D model to be presented in the small FOV of the simulator’s display. This may be considered to be scene compression since the viewer is seeing a wide FOV in a display which has a smaller FOV. Below, we report two studies of optic flow in our driving simulator when various sizes of GFOVs were used.
Methodology ______________________________________________________________________________________ ____
DSC 2007 North America – Iowa City – September 2007
Apparatus Northeastern University’s Virtual Environments Driving Simulator, shown in Figure 1, was used to collect data.
Figure 1. Driving simulator configuration. An LCD projector mounted behind the vehicle, displayed a 45 degree horizontal by 33.75 degree vertical image on a curved screen in front of the vehicle. The resolution was 1600 x 1200 pixels and the frame rate was constant at 60 frames per second. The vehicle’s force feedback steering wheel, and accelerator and brake pedals were connected to the computer, which rendered the scene according to inputs from these devices. The speedometer readout was covered for the duration of the experiment. Participants (Study 1) Twenty subjects (10 female and 10 male) participated in the study. They ranged in age from 22 to 35 years of age and had 20/20 or corrected to 20/20 vision. All subjects held a valid driver’s license and had driven for at least 3 years. Procedures (Study 1) Participants were given a short test run in the simulator to become familiar with the steering, acceleration, and braking characteristics of the vehicle. Each participant was requested to make paired comparisons. For consistency and clarity, subjects were read the following instructions, “In each of the following trials, you will be shown scene 1 which consists of a straight roadway with approaching vehicle followed immediately by a scene 2 with approaching vehicles. You will be asked if the speed of the cars in the second presentation was slower, faster, or the same as the speed of the cars in the first presentation.”
Experimental Design (Study1) ______________________________________________________________________________________ ____
DSC 2007 North America – Iowa City – September 2007
Four values of GFOV (25, 45, 65, and 85 degrees) were used for this study. Subjects were tested on 32 paried comparisons delivered in two trials of 16 randomly ordered pairs. In Figure 2 is the 3D experimental scene used for study 1.
Figure 2. Vehicles approaching at 30 mph. Participants (Study 2) Thirty subjects (15 female and 15 male) participated in the study. They ranged in age from 22 to 35 years of age and had 20/20 or corrected to 20/20 vision. All subjects held a valid driver’s license and had driven for at least 3 years. Each participant wea paid a $10 honorarium for participating in the study. Procedures (Study 2) Participants were given a short test run in the simulator to become familiar with the steering, acceleration, and braking characteristics of the vehicle. Each participant was requested to produce vehicle velocities. For consistency and clarity, subjects were read the following instructions, “In each of the following runs, you will see a straight roadway. Before you begin each run, I will tell you a target speed, which will either be 30 miles per hour or 60 miles per hour. As soon as you feel that you are traveling at the target speed, pull up on the turn signal to the left of the steering wheel. There will be twelve runs, a short break, and then twelve more runs”. Each time the driver activated the turn signal, the computer recorded the vehicle’s instantaneous velocity. Experimental Design (Study 2) For this research, a 2 x 2 x 3 experimental design was used. Factor A was Target Velocity (30 or 60 mph). Factor B was Optic Flow (low or high). Factor C was Geometric Field of View (25, 55 or 85 degrees). Each participant was tested on the 12 conditions twice. The order of the presented conditions was randomized across runs. In Figure 3 is the high optic flow condition. The trees were not present in the low optic condition.
______________________________________________________________________________________ ____
DSC 2007 North America – Iowa City – September 2007
Figure 3. High optic flow condition.
Results (Study 1) For each paired comparison the subjects were asked if the velocity of the vehicles in the second presentation appeared to be less, equal, or greater than the velocity of the vehicles in the first presentation. Figure 4 presents the percentages that they reported equal velocities when at least one of the paired comparisons was an 85 degree FOV.
Figure 4. Paired Comparisons with an 85 degree GFOV. Note that when the paired comparison involved exactly the same conditions (a GFOV of 85 degrees followed by a GFOV of 85 degrees) subjects correctly reported this 85.5 percent of the time. However, when one of the conditions consisted of another GFOV (65, 45, or 25) paired with the 85 degrees GFOV, the percent of equal comparisons ______________________________________________________________________________________ ____
DSC 2007 North America – Iowa City – September 2007
reported were 45, 12, and 1 respectively. These differences were the result of subjects perceiving that the vehicles in the 85 degrees GFOV presentation were going faster then in the other presentations. Unknown to the subjects was the fact that the velocity of the oncoming vehicles was 30 mph for all GFOVs. Thus the perceived velocity of the autonomous vehicles was highly dependent on the geometric field of view used in connection with the driving simulator. Subjects were also able to discriminate the perception of vehicle velocity when the difference in the GFOVs in a paired comparison trial was 20 degrees as shown in Figure 5 below. Paired comparisons involving 25-45 or 45-25 resulted in equal responses only 17.5 percent of the time. Paired comparisons involving 45-65 or 65-45 resulted in equal responses only 21.5 percent of the time. While paired comparisons involving 65-85 or 85-65 resulted in equal judgments 45 percent of the time.
Figure 5. Percent judged equal when the GFOV was different by 20 degrees. The results of study 1 clearly established that subjects perception of vehicle velocity was dependent of the size of the geometric field of view.
Results (Study 2) In study 2, subjects were asked to produce a target velocity of 30 or 60 mph. A three-factor ANOVA was used to analyze the results. Tukey post hoc tests were used to establish the degree of pairwise significance of means for F ratios at p < .05.
______________________________________________________________________________________ ____
DSC 2007 North America – Iowa City – September 2007
All three main effects, Target Velocity (F=317.4, df=1,708), Optic Flow (F=12.3, df=1,708), and Geometric Field of View (F=247.1, df=2,708) were highly significant with p < .001 and are presented in Figures 6, 7, and 8, respectively. A surprising finding was the amount of error resulting when participants attempted to produce a velocity of 30 mph. Here, the magnitude of error was nearly 20 mph over the target velocity as shown in Figure 6. Mean Mean Velocity Velocity
Produced Velocity Velocity (mph) (mph) Produced
70 70 60 60 50 50 40 40 30 30 20 20 10 10 00
30 30
60 60
Target Target Velocity Velocity (mph) (mph)
Figure 6. Produced velocity by target velocity. Figure 7 below shows that the presence of high optic flow resulted in the mean produced velocity being less. However, the amount of difference between low and high optical flow was only 2.4 mph. The presence of a dashed lane divider in both the low and high optic flow conditions may have contributed to this small difference. Mean Mean Velocity Velocity
Produced Velocity Velocity (mph) (mph) Produced
70 70 60 60 50 50 40 40 30 30 20 20 10 10 00
lo loww
high high
Amount Amount of of Optical Optical Flow Flow
Figure 7. Produced velocity by optical flow. As seen below, Figure 8 illustrates that mean velocity production consistently decreased as the GFOV increased; this suggests that a wider geometric field of view leads to the perception of traveling at a higher velocity, yielding lower velocity production. ______________________________________________________________________________________ ____
DSC 2007 North America – Iowa City – September 2007
Mean Mean VVelo elocity city
Produced Velocity Velocity(mph) (mph) Produced
7700 6600 5500 4400 3300 2200 1100 00
2255
5555
8855
GGeo eometric metric Field Field ooff VView iew
Figure 8. Produced velocity by GFOV. A highly significant interaction was found (p < .001) between the target velocity and GFOV as shown in Figure 9. This interaction can be largely attributed to participants producing very high velocities for the GFOV of 25 degrees when the Target Velocity was 30 mph. Mean Mean Velocity Velocity mp mphh 30 30
mp mphh 60 60
Produced Velocity Velocity (mph) (mph) Produced
70 70 60 60 50 50 40 40 30 30 20 20 10 10 00
25 25
55 55
85 85
Geometric Geometric Field Field of of View View
Figure 9. Interaction of produced velocity with GFOV. When the GFOV was 85 degrees, subjects’ production for the 60 mph target velocity was fairly acdurate (56.5 mph). However, when asked to produce a target velocity of 30 phm, they produced an average velocity of 38.5 mph.
Discussion The finding that mean velocity production decreased as the GFOV got larger was expected. This confirms the results of Adetiloye et. al. [2005] that participants perceive ______________________________________________________________________________________ ____
DSC 2007 North America – Iowa City – September 2007
oncoming vehicle velocities to be faster when viewing an 85 degree GFOV as compared with smaller GFOVs. If drivers feel that they are going faster at higher GFOVs, then production of velocity at these GFOVs would consequently be less. The large amount of overproduction error (20 mph) found when subjects tried to produce a velocity of 30 mph has implications for all simulators that change the viewer’s position at low rates, i.e. walking and driving simulators. This result suggests that the production of a 30 mph velocity (and presumably slower velocities) in a simulator is very different from that in the real world. Triggs and Berenyi [1982] reported on participants estimating vehicle velocities in the real world. When trying to estimate a velocity of 33 mph, the error during nighttime driving was only 1 mph. Under daytime driving, the error was 6 mph less than the actual speed of 33 mph. Casual observations of drivers in neighborhoods (where speed limits are 25/30 mph) found that drivers can accurately produce velocities in the range of 25 to 35 mph, and realize when they are driving too fast. Subjects failed to produce a target velocity of 30 mph, regardless of the size of the GFOV was surprising. This may be an inherent problem associated with driving simulators. It causes a conflict with what is being reported by the vehicle’s speedometer. The speedometer may be reporting 40 mph but a subject’s perception of the virtual environment indicates that the vehicle is going 30 mph. This phenomenon needs further study. The finding of little difference (2.5 mph) between the low (no trees) and high (with trees) optical flow conditions may be due to both conditions having lane markings. Under the low optical flow condition, participants may have used the flow of the lane markings to produce the requested velocity. The addition of the trees to that environment added only a nominal amount to their perception of the amount of optical flow.
______________________________________________________________________________________ ____
DSC 2007 North America – Iowa City – September 2007
References Adetiloye, C., Wu, Q., & Mourant, R. 2005. Perception of optical flow and geometric field of view. (poster) Proceedings the 32nd International Conference on Computer Graphics and Interactive Techniques (SIGGRAPH 2005). Asano, Y. & Uchida, N. 2005. Improvement of Driver’s Feeling by Turning Cabin Driving Simulator. Proceeding of the Simulator Conference – North America, Orlando, Florida, 230-239. Evans, L. 1970. Speed estimates from a moving vehicle. Ergonomics, 13, 219-230. Gibson, J.J. 1979. The Ecological Approach to Visual Perception. Houghton Mifflin: Boston, MA. Lee, D.N., 1980. The optic flow field: the foundation of vision. Phil. Trans. Royal Society London B 290, 169-179. Triggs, T. J.,& Berenyi, J.S. 1982. Estimation of automobile speed under day and night conditions. Human Factors, 24(1), 111-114.
______________________________________________________________________________________ ____
