Graphics Reference
In-Depth Information
from country to country across shape, color, text font, and language. For detection of
circular road signs Zelinski Transform [
26
] gives good results, while an effective way
to capture general traffic shape information is histogram of gradients (HOG) [
27
].
3.4.4.2 Estimating Depth and Motion and Pedestrian Detection
Pedestrians are the most vulnerable participants in traffic and vision-based recogni-
tion of pedestrians is a key problem of smart vehicles. Pedestrian detection requires a
variety of features (i.e., Haar wavelets, HOG, CoHOG, LRF, etc.) and classification
methods (i.e., Adaboost, SVM). A number of related surveys exist on pedestrian
detection [
6
,
28
-
30
,
39
,
40
].
In most cases, precise and robust estimation of depth and motion is a key ele-
ment for pedestrian detection and obstacle avoidance functions in automotive driver
assistance [
31
]. In order to maximize knowledge about the scene ahead, every sensor
pixel should be utilized. By calculating position and motion of every frame pixel,
the front view camera system can anticipate more accurately pedestrian behavior.
Therefore, it is very beneficial to obtain dense disparity and dense optical flow. On
the other hand, optical flow and stereo vision are extremely challenging algorithms
for embedded environment due to their compute and bandwidth requirements.
The perception of motion is instrumental for ADAS. The optical flow algorithm
is a key building block for ADAS functions such as estimation of vehicle egomotion,
obstacle/pedestrian detection, detection of cars in the blind spot and structure from
motion. Optical flow used in advanced driver assistance environment must cope
with weakly textured areas, large displacements, and constant change of illumination
conditions caused by either scene change or unanticipated and unknown adjustments
of camera parameters such as exposure time.
Most variation-based optical flow algorithms use the brightness constancy con-
straint between successive frames, which is often violated in real-world scenes while
driving. Second serious limitation of typical optical flow methods suggested in liter-
ature is that they can only cope with small displacements, while typical real-world
scenes exhibit large motion vectors especially in curves or areas near the vehicle. A
quantitatively different approach was suggested in [
32
], where the Census Transform
was used to represent small image patches and the primitives were matched using
a table based indexing scheme. Another approach to improve robustness to change
in brightness is to replace the simple gray value illumination constancy constraint
from classical optical flow approaches by Hamming distances between different
census signatures [
33
]. To overcome large displacement challenge, integration of
rich descriptors into the variational optical flow setting is suggested in [
34
]. Use
of Rudin-Osher-Fatemi (ROF) denoising scheme [
35
] to achieve resistance to illu-
mination change is proposed by several authors [
36
]. While the method improves
resistance to illumination change it is computationally intensive and struggles with
large displacements evenwhen a pyramid scheme is used. Adifferent approachwhich
explicitly models the varying illumination by adding an additional scalar function is
proposed in [
37
]. The function is estimated in a joint optimization of the optical flow
