IEEE Robotics & Automation Magazine - June 2014 - 28

Missed Tree Missed Lamp Posts
20

0

Detected Tree
20

0

Detected Lamp Posts
(with Spatial Errors)

False Alarm

False Alarms

-20

0

40

20
(a)

0

20
(b)

40

8

8

0

-20

0

extract line segments from the
data [10]. Due to RANSAC's
random nature in selecting
its  initial model inlier points,
Figure 2(c) shows the line correctly detected, whereas Figure  2(d) shows that the true
wall is missed and a false line is
detected. Similar to the radar
detection concepts, a possible  detection threshold for
RANSAC can be set, again
yielding an approximate Pfa
and, depending on the number
of data inliers within the actual
line, PD .

Visual Feature Detection
with SURF
In the case of the camera, a
-8
-8
well-known blob detector called
-8
0
the SURF detector is applied in
0
8
8
-8
(c)
(d)
Figure 2(e) and (f). This detector searches for local maxima in
the determinant of Hessian
(DoH) applied to a two-dimensional (2-D) array of image pixels at various scales [11]. Such a
maxima is an indication of a
distinctive region in the image
space corresponding to a possible feature likely to be detected
from other locations. The SURF
(e)
(f)
descriptors are shown as circles,
the area of which correspond to
Figure 2. The random nature of detections, based on a radar CFAR detector, line-based RANSAC, and
the scale at which SURF deterthe SURF detection algorithm: (a) and (b) radar point detections based on an OS−CFAR detector, (c)
mined the maximum DoH, at
and (d) laser range data with a single RANSAC line detector, and (e) and (f) visual SURF features.
the pixel center of that circle. In
Figure 2(f) the ambient light
detector applied to the received power values [8], [9]. All level has reduced slightly. In low-light environments, the campower values exceeding the OS-CFAR threshold are shown as era increases its gain, which also amplifies the noise. In any
crosses, together with the superimposed ground truth loca- blob detector (e.g., SURF), this variation in pixel intensities intion of objects (trees, lamp posts, and so forth) shown as troduces randomness into the detection process. The green
ellipses. The two experiments were carried out in the same circles in Figure 2 correspond to SURF features, which were
environment under the same environmental conditions. It is detected in over 90% of the 300 recorded images of the same
evident that, due to the random power noise in each radar scene, and are therefore taken to approximate ground truth
scan, some of the features detected in one image are missed in SURF features. In this case, varying the light conditions in a
the other, as is the random presence of false alarms. Under an scene so that the DoH varies can give approximate but meanOS-CFAR detector, the probabilities of false alarms Pfa and ingful Pfa and PD values for the SURF features.
detection PD can be quantified under certain noise distribution assumptions.
General Sensing Errors
Although spatial errors occur in terms of the estimated
RANSAC Line Detection with LRFs
location of the features, of far greater importance is the fact
Figure 2(c) and (d) shows data samples from a 360° laser that the correctness of the detections themselves is quesrange scan recorded in an environment in which a single wall tionable. Consequently, the realistic uncertainty associated
was present (labeled). The RANSAC algorithm was applied to with measurement data includes detection (or feature
28

*

Actual Wall
(Detected)
Detected Line

IEEE ROBOTICS & AUTOMATION MAGAZINE

*

June 2014

Detected Line
Actual Wall (False)
(Missed)
Table of Contents for the Digital Edition of IEEE Robotics & Automation Magazine - June 2014
