02-03-2013, 11:58 AM
Context Enhancement of Nighttime Surveillance by Image Fusion
Context Enhancement.pdf (Size: 286.73 KB / Downloads: 26)
Abstract
In this paper, we propose a novel method of automatically
combining images of a scene at different time intervals
by image fusion. All the important information of the
original low quality nighttime images is combined with the
context from a high quality image of the daytime at the same
viewpoint. The fused image contains a comprehensive description
of the scene which is more useful for human visual
and machine perception. Experimental results show that the
proposed method is robust and effective.
Introduction
Visual surveillance as an active topic in computer vision
has received much attention in recent years. Visual surveillance
in dynamic scenes aims to detect, recognize and track
objects such as people and cars from image sequences at
low level. Moreover, it analyses and understands objects’
behavior at a high level.
However, most of the previous work has focused on daytime
surveillance. Understanding nighttime video is a challenging
problem because of the following reasons: Firstly,
due to low contrast, we can not clearly extract moving objects
from the dark background. Most color-based methods
will fail on this matter if the color of the moving objects
and that of the background are similar. Secondly, the signal
to noise ratio is usually very low due to high ISO (ISO is
the number indicating camera sensors sensitivity to light).
Using a high ISO number can produce visible noise in digital
photos. But low ISO number means less sensitivity to
light. Thirdly, environment information, in other words, the
context information of a scene, affects the way people perceive
and understand what has happened. Due to the limited
information in nighttime video, understanding behavior of
people in video becomes more difficult.
Enhancement of Nighttime Video
As we mentioned in Section 1, due to the low contrast
of the nighttime video, it becomes more difficult to accurately
extract moving objects from the dark background.
How to improve the contrast of nighttime video while suppressing
the noise becomes a crucial problem. Bennett
et al. [1] present an adaptive Spatio-Temporal Accumulation
(ASTA) filter for reducing noise in low dynamic range
(LDR) videos. In addition, a tone mapping function is proposed
to enhance LDR videos. The whole process takes
approximately one minute per frame of size 640 × 480. In
this paper, we demonstrate a successful application of this
tone mapping function to the nighttime video enhancement.
Motion Detection
In this section, mixed Gaussian model [2] is adopted to
motion detection. Mixed Gaussian model is an effective solution
to real time motion detection due to its self learning
capacity. In addition, this method is robust to variations
in lighting, moving scene clutter, multiple moving objects
compared with other methods. The mixed Gaussian model
for real time motion detection can be briefly summarized as
follows [2]:
(1) Every new pixel value is checked against the existing K
Gaussian distributions. A match is defined as a pixel value
within 2.5 standard deviations of a distribution.
(2) Sort Gaussians and determine if the Gaussian is background.
(3) Adjust the prior weights of the distributions.
(4) Adjust the mean and standard deviation for matched distributions.
Conclusions
In this paper, we have presented a novel approach to
enhancing the context of nighttime surveillance by image
fusion. The proposed method provides a real time and
robust solution to front-end image pre-processing in nighttime
video surveillance. In addition, our method is flexible
in adapting to different scenarios. The resultant image
contains a more accurate and comprehensive description
of the scene which is more useful for human visual and
machine perception.