Perceived collision with an obstacle in a virtual environment Russell L Woods, Jennifer C Shieh Laurel Bobrow, Avni Vora, James Barabas, Robert B Goldstein and Eli Peli Schepens Eye Research Institute and Harvard Medical School, Boston, MA ARVO 2003
How do you define a collision? In the literature: Center to center No consideration of the physical size of the observer or “safe distance” Evaluated visual information (e.g. , TTC, heading perception) or cognitive issues (e.g. search) Way-finding
Cutting, Vishton & Braren (1995) Collision detection from relative motion of obstacle and other objects Stick figures, sparse environment Simulated fixation task Center to center Angular perspective Cutting, Vishton & Braren (1995)
Are potential-collision decisions based on physical size? (i.e. how big you are)
The task Walk on a treadmill (self propelled) Rear projected screen (77 cm, 95 degrees wide) “infinite” shopping mall corridor Obstacle appeared at 5m or 15m for 1 second Square pillars with images of people (30cm or 70cm wide) Task: Would you have collided with the obstacle? New path before each obstacle Random angular offsets of paths
Closest distance to obstacle
Some obstacles crossed the path
Obstacle appearance distance
The task Walk on a treadmill (self propelled) Rear projected screen (77 cm, 85 degrees wide) “infinite” shopping mall corridor Obstacle appeared at 5m or 15m for 1 second Square pillars with images of people (30cm or 70cm wide) Task: Would you have collided with the obstacle? New path before each obstacle Random angular offsets of paths
small obstacle, 5m, 55cm
large obstacle, 15m, 100cm
The task Walk on a treadmill (self propelled) Rear projected screen (77 cm, 85 degrees wide) “infinite” shopping mall corridor Obstacle appeared at 5m or 15m for 1 second Square pillars with images of people (30cm or 70cm wide) Task: Would you have collided with the obstacle? New path before each obstacle Random angular offsets of paths
“Yes, collision” responses against closest distance to obstacle
How “big” do you feel? Distance with optimal decision (highest kappa) Kappa coefficient of association How “good” a decision? Decision quality = maximum kappa (height) How “big” do you feel? Distance with optimal decision (highest kappa)
Collision envelope varied between subjects and with obstacle distance No effect of obstacle size Z19=3.44 p<0.001 22 subjects
Better decisions at smaller obstacle distance Z19=4.07 p<0.0001 22 subjects
Some subjects had great difficulty at 15m
Do physical characteristics matter? Preferred walking speed, stride length Width at shoulder and of the arms Age Height, weight, body mass index (BMI)
Collision envelope was not predicted by physical characteristics 5m rs = 0.02, p=0.92 15m rs = 0.01, p=0.99 22 subjects
Collision envelope was (usually) larger than measured physical characteristics Collision envelope equals body width 5m rs = -0.26, p=0.25 15m rs = +0.03, p=0.92 22 subjects
Further experiments Repeatability 15m, was task difficulty due to poor determination of heading? Does physical size not matter at all?
How repeatable were our results? 5m rs = 0.43, p=0.26 15m rs = 0.77, p=0.08 Compare distributions No significant differences (p>0.69) 8 subjects
15m obstacles: was task difficulty due to a problem determining heading?
15m obstacles: providing heading information improved task performance z4 = 1.15, p = 0.25 z4 = 2.37, p = 0.02 5 subjects
Does physical size not matter at all? Wings Does physical size not matter at all?
Does physical size matter? Yes Actual (half) width of the wings z4 = 2.02, p = 0.04 z4 = 1.83, p = 0.07 5 subjects
Review of main results Effect of distance collision envelope slightly larger; and decision quality reduced at further distance Heading perception seems a limiting factor Physical characteristics not predictive, but Collision envelope can be manipulated
We evaluated …. While…. But … Collision detection Subject’s perception of “size” (collision envelope or safety margin) While…. Free viewing in “rich” virtual environment Actually walking But … Stationary obstacles only Single obstacles only
Thank you (for coming to the last presentation at ARVO 2003) Supported by NIH grant EY12890
The collision envelope We defined the collision envelope as the optimal decision point of the intra-class kappa coefficient This assumes that the cost of a false positive (avoidance when no risk) is the same as a false negative (collision)
The weighted kappa coefficient K0 The weighted kappa coefficient K0.1 places greater cost on false negative (collision) K0.1 K0.5 5m 46cm 37cm 15m 65cm
On a treadmill
large obstacle, 5m, 25cm
large obstacle, 15m, -15cm