Graphics Reference
In-Depth Information
3.4.4 Vision Building Blocks
3.4.4.1 Common Preprocessing and Feature Detection
Computer vision has entered the mainstream of ADAS. Computer vision algorithms
detect road lanes, objects, and pedestrians, anticipate and augment driver's actions.
There are no industry standards for embedded vision processing algorithms (unlike
for video codecs used for video compression) and if we ask ten different people to
develop a forward collision warning, we will end up with ten different solutions.
To properly capture images of moving objects, exposure time and shutter type
are critical. Concurrent algorithms/functions running in the system typically require
different exposure settings and some functions such as TSR require more than one
exposure setting. To accommodate these requirements typical ADAS imaging sensors
have high dynamic range and output multiple exposures.
Some common preprocessing steps in the driver assistance systems based on
the embedded vision are: custom image signal processing (ISP); various filtering,
creation of image pyramids; integral image; and feature extraction. The ISP block
is typically taking a raw high dynamic range input from an imaging sensor and
converting to a format used by the algorithms.
Surround view and rearview systems rely on cameras with wide field of view
at a cost of large distortion. Lens distortion correction is typically one of the first
processing steps in these systems. Depending on situation, the software can change
viewing angle (i.e., render top view image) before displaying video to the driver. The
most efficient way to handle this homographic transform is via look-up table.
Surround view systems combine inputs frommultiple satellite cameras positioned
around car in one 360-degree image (Figs.
3.1
,
3.4
, and
3.5
). In computer vision,
extensive work has been done in the field of image stitching and photogrammet-
ric matching. During image registration features (using feature descriptors such as
SIFT, ORB, SURF [
54
] etc.) are extracted from each view, and matched to fea-
tures in the other overlapping views. Random sampling consensus (RANSAC) of
matched features is used in order to remove any mismatched points and to estimate
the homography between overlapping views.
Key components of lane detection are: extraction of road markings; road model-
ing; post processing; and position tracking. Edge-based techniques for road marking
extraction give good results on solid and segmented lines, but they typically fail in
situations which contain many nonessential lines. Different approaches have been
proposed to overcome this limitation [
25
]. To improve estimates, it is necessary to
include a priori knowledge of the road in the postprocessing steps. The most com-
mon tracking technique used in lane-position-detection systems is Kalman filtering.
A survey of vision-based lane detection can be found in [
25
].
In some systems, color is required for traffic sign recognition algorithms, while
in most cases vision-based ADAS functions rely purely on monochrome imaging
sensors as they offer higher light sensitivity compared to color sensors. Traffic sign
recognition is a classic example of rigid object detection although road signs differ
