Information Technology Reference
In-Depth Information
Fig. 13.4
Illustration of the Farneback dense optical flow vectors for sample video frames of
different traffic congestion levels in the UCSD dataset—
a
light traffic
b
medium traffic
c
and heavy
traffic
modeling
,
frame differencing
and
optical flow
. Background modeling is more suitable
to model slowly changing backgrounds. Frame differencing, on the other hand fails
in the situations with slowly moving objects and produces many “holes” in detected
region(s), if object(s) exhibit poor texture. We decided to employ optical flow meth-
ods for the detection of moving object(s). Optical flow is the apparent motion of
brightness patterns in images caused by the relative motion between an observer and
the scene. The main concern about optical flow is its time-consuming computation.
However, recent works claim that optical flow can be computed fast and accurately.
We explored state-of-the-art sparse and dense optical flow extraction methods to
detect moving objects. In our final implementation, in order to extract dense optical
flow vectors we decided to use the Farneback method [
8
] due to its superior clas-
sification performance. The optical flow vectors are low-level features and provide
a rough overview of the traffic flow in the scene. Therefore, they are processed to
construct meaningful motion descriptors (i.e., mid-level features) based on their sta-
tistics. In Fig.
13.4
, sample video frames of different traffic levels are shown with the
corresponding dense optical flow vectors.
13.3.2.3 Traffic Flow Analysis Model
As the final step of VA, we train a multiclass SVM in order to learn a traffic flow
analysis model. This model is trained to classify traffic flow extracted from videos
into one of the following three categories:
heavy
,
medium
,or
light
. Training of the
model was performed using the UCSD dataset which is introduced in Sect.
13.3.2.4
.
13.3.2.4 Video Dataset and Ground Truth
We currently use the UCSD highway traffic dataset which contains 254 highway
traffic videos of daytime highway traffic in Seattle (Washington, USA) in order
to build our traffic analysis model. The videos are of 5 s length with 320
240
resolution recorded at 10 frames per second (fps) and collected from a single static
×
Search WWH ::
Custom Search