Specifications

7
objects of interest. It will only guarantee to find images without moving objects. The automatic
algorithm will select scenes where vehicles are temporarily stopped. However, the user can use
the automatic algorithm and then proceed by manually selecting a different background image if
the automatic selection is inappropriate.
Each image is converted to gray scale prior to point-wise comparison between the background
image (B) and the scene image (I). The result is an intensity difference image (D). Pixels with
intensity greater than the object detection threshold θ are set to a value of 1. The rest of the
pixels are set to a value of 0. The result of this operation is an object mask image (M) that
specifies the candidate locations of objects. The threshold θ can be adjusted according to the
image noise levels of the camera in order to minimize the number of false alarms. A high θ can
also be utilized to combat false alarms due to slight movement of objects in the background. For
example, wind can cause trees or signs to move slightly.
To further reduce the sensitivity of object detection process to false alarms, a series of
morphological operations are performed on the object mask. These operations include eroding
the mask through a majority operation in order to eliminate regions that are above threshold but
too small to be an object of interest, dilating to assure that the entire object is captured within the
extent of the mask, and finally closing to create convex regions that have no holes. Given the
complexity of the point-wise comparisons and morphological operations, intensity images are
down sampled by a factor of Δ prior to processing.
Once the object mask has been computed, separate regions are identified and labeled. A
bounding box, center of mass, and area is then computed for each. Objects with an area less than
a minimum area threshold ϕ are discarded. Objects with an area between ϕ and the minimum
vehicle area threshold ν are labeled pedestrians. Objects with an area greater than or equal to ν
are labeled vehicles.
Because the light levels in the scene change with time and the object detection algorithm
depends on image intensity, any given background image will only be appropriate for use with
images that were taken at a point nearby in time. To address this problem, the background is
adapted slowly over time. For each frame f, a background image for the next frame B(f+1) is
created based on the current background image B(f), the object mask M(f), and regions of the
current image I(f) that contain no objects M
i
(f), such that
B(f+1) = B(f) * β + [ B(f) • M(f) + I(f) • M
i
(f) ] * (1-β),
where β is a parameter that controls the time constant of adaptation to changes in the scene, *
denotes scalar image multiplication, and • denotes point-wise image multiplication. Using this
approach, the background image is updated using the current image locations that do not contain
objects.