10-12-2012, 12:08 PM
Night vision pedestrian detection using a forwardlooking infrared camera
Night vision pedestrian detection.pdf (Size: 692.08 KB / Downloads: 43)
Abstract
This paper describes a night vision pedestrian
detection system for autonomous vehicles using an onboard
forward-looking infrared (FLIR) camera. Much emphasis has
been placed on real-time processing requirements and robustness
of the system under different scenarios. At low level, the Haarlike
features are used to discriminate infrared pedestrians.
AdaBoost learning algorithm is employed to select the most class
relevant infrared pedestrian features. For effective sub-windows
scanning, a novel keypoint based region of interest (ROI)
selection strategy in IR imageries is proposed. The system has
been implemented as part of a preliminary real-time driver
assistance system of Vision Miracle Intelligence Company,
Changsha, China. Experimental results under different urban
scenarios prove that the proposed system is robust and efficient.
INTRODUCTION
Pedestrian detection is an important research field for a
number of applications such as smart video surveillance, driver
assistance system, content based image/video indexing,
robotics and military applications. Driven by the decreasing
cost of infrared (IR) sensors, night vision systems have gained
more and more interest in recent years, and thus increasing the
need of pedestrian detection at night [1-4].
A considerable amount of previous work has addressed the
problem of vision based pedestrian detection, using visual or IR
imagery in both monocular and stereo-camera configurations,
from transportation surveillance infrastructure as well as
moving vehicles. For more recent comprehensive reviews of
the literature, we refer the reader to the work of Gandhi and
Trivedi [5] and the work of Enzweiler and Gavrila [6].
Pedestrian detection from moving platforms is a challenging
task because of a wide range of possible pedestrian
appearances and poses, different pedestrian motion patterns,
uncontrolled outdoor environments, moving camera, and other
stationary as well as independently moving objects [3, 5, 6, 7,
8]. The complexity of the problem is augmented by stringent
performance criteria and hard real-time constraints.
FEATURE EXTRACTION AND SELECTION
Recently the boosting-based detector proposed by Viola
and Jones [9] receive lots of attention in face detection research.
In this paper, we adopt a similar idea for pedestrian detection in
IR imageries. Real-time performance is achieved by learning a
sequence of simple Haar-like rectangle features. The Haar-like
features encode differences in average intensities between two
rectangular regions, and they can be calculated rapidly through
integral image. The complete Haar-like feature set is large and
we use AdaBoosting algorithm to select a small number of
distinctive rectangle features and construct a powerful classifier.
Haar-like rectangle features measures the difference
between the average intensities of rectangular regions. Each
feature is composed of two or three ‘black’ and ‘white’
rectangles joined together. The feature value is calculated as a
weighted sum of two components: the pixel gray level values
sum over the black rectangle and the sum over the whole
feature area. The weights of these two components are of
opposite signs and for normalization purpose. Their absolute
values are inversely proportional to the areas. Direct
computation of pixel sums over multiple rectangles would be
very slow and not suitable for real-time pedestrian detection.
CONCLUSION
A night vision pedestrian detection system for autonomous
vehicles is presented in this work, using an onboard forwardlooking
infrared (FLIR) camera. Firstly, Haar-like features are
extracted to discriminate infrared pedestrians. Secondly,
AdaBoost learning algorithm is employed to select the most
relevant infrared pedestrian-specific features. Finally, a
keypoint based sliding window cascade classifier is utilized for
effective pedestrian localization. Experimental results under
different urban scenarios demonstrate the robustness and
effectiveness of the proposed system.