Classification Performance

test hand arm

position

(X,Y) absolute 5.8 pixels 22.0 pixels

(0.9% rel) (3.4% rel)

(X,Y) residual 0.7 pixels 2.1 pixels

(0.1% rel) (3.2% rel)

orientation

() residual 4.8 degrees 3.0 degrees

(5.2% rel) (3.1% rel)

Table 4.1:
Absolute position error is the RMS error between the reported position and the guide path. Residual position error the RMS error after model errors accounted for. Orientation error is the RMS error to a low-passed filtered version of the data since no ground truth was available. ``rel'' designates relative error: the ratio of the error to the total path length.

**Table 4.1:**
Absolute position error is the RMS error between the reported position and the guide path. Residual position error the RMS error after model errors accounted for. Orientation error is the RMS error to a low-passed filtered version of the data since no ground truth was available. ``rel'' designates relative error: the ratio of the error to the total path length.
test	hand	arm
position
	(X,Y) absolute	5.8 pixels	22.0 pixels
		(0.9% rel)	(3.4% rel)
	(X,Y) residual	0.7 pixels	2.1 pixels
		(0.1% rel)	(3.2% rel)
orientation
	() residual	4.8 degrees	3.0 degrees
		(5.2% rel)	(3.1% rel)

Pfinder usually obtains video signal from a JVC-1280C, single CCD, color camera. It provides an S-video signal to the SGI digitizers. Its specs claim 40db signal to noise. The SGI digitizer smoothes and sub-samples the images down to sixteenth resolution. Give this input, Pfinder exhibits Root Mean Squared (RMS) tracking errors on the order of a few pixels in position and a few degrees in orientation, as shown in Table 4.1. Here, the term ``hand'' refers to the region from approximately the wrist to the fingers. An ``arm'' extends from the elbow to the fingers.

These figures show example data from of independent runs of tracking while the hand was slid along a straight guide. (a) is the result of tracking the left hand. (b) is the result of tracking the right hand. (c) is the result of tracking the left arm. (d) is the result of tracking the right arm. Hand refers to the visible flesh area when a person wears a long sleeve shirt. Arm refers to the visible flesh when a person wears a short sleeve shirt.

For the translation tests, the user moves through the environment while holding onto a straight guide. Absolute errors (less then 6 pixels for the hand, and 22 pixels for the arm) are measured from the reported position to the position of the guide path in the image. Residual errors (sub-pixel for the hand, and little more than 2 pixels for the arm) ignore errors intrinsic to the experimental setup: the tracked region is constrained to move in a straight line in the image, but nothing constrains the center of the tracked area to be coincident with the center of the guide. Note that these accuracies are obtained from images that are sub-sampled by a factor of four in each dimension. Relative error is the ratio of the RMS error to the total screen size. Figure 4.1 shows some sample results from the trials.

All plots if Figure 4.1 show reported X position versus reported Y position. Plot 4.1a shows data from tracking the left hand of a long-sleeved subject, so the tracked region extended from the wrist to the fingers. Plot 4.1b shows data from a similar run on the subject's right hand. Plot 4.1c shows data from tracking the left arm of a short-sleeved subject. In this case the tracked region extended from approximately the elbow to the fingers. Plot 4.1d shows data from a similar run on the subject's right arm. This run exhibits some instability, probably due to a combination of factors including transient partial occlusions of the arm.

For the orientation error test, the user smoothly moves an appendage through several cycles of approximately 90 degree rotation. There is no guide in this test, so neither the path of the rotation, nor even its absolute extent, can be used to directly measure error. The RMS distance to a low-pass filtered version of the data provides a measure of the noise in the data. Figure 4.2 shows some data from the orientation test.

All plots if Figure 4.2 show reported

position versus time. Plot 4.2a shows data from tracking the left hand of a long-sleeved subject, again the tracked region extended from the wrist to the fingers. Plot 4.2b shows data from a similar run on the subject's right hand. Plot 4.2c shows data from tracking the left arm of a short-sleeved subject. As in the position tests the tracked arm region extended from the elbow to the fingers. Plot 4.2d shows data from a similar run on the subject's right arm. The hand orientation data exhibits more noise. This likely due to the fact that the tracked region in the arm case has significantly more eccentricity than in the hand case. That extra information allows the model blobs to be more stable in orientation.

These figures show the results of independent runs of an experiment designed to test stability of orientation tracking. No guide was used, the subject attempted to move the specified appendage smoothly through several iterations of 90 degrees. (a) is the result of tracking the left hand. (b) is the result of tracking the right hand. (c) is the result of tracking the left arm. (d) is the result of tracking the right arm. Hand refers to the visible flesh area when a person wears a long sleeve shirt. Arm refers to the visible flesh when a person wears a short sleeve shirt.