Information Technology Reference
In-Depth Information
congestion (light, medium, or heavy). The rationale behind the use of density and
speed is that heavy congestion results from high density and low speed, while light
congestion is synonym of lower density and higher speed. However, the authors have
not provided any information regarding the applicability of their method to scenes
showing more than one traffic direction, and the results only illustrate one-way traffic
situations.
A similar solution based on density and speed has been developed by Hu et al. [ 10 ].
The lane where vehicles are present is determined by aggregating the location of the
moving blobs over time. A variant of the KLT [ 15 ] is used to track corner points
extracted in the regions covered by moving blobs. Classification is achieved via
Fuzzy Logic. Similarly to [ 24 ], the authors only demonstrated this approach on one-
way situations.
In an attempt to develop an alternative to optical flow analysis, Derpanis and
Wildes [ 6 ] have suggested using a holistic solution derived from the field of image
texture analysis. The authors represent image dynamics using spatiotemporal orien-
tation decomposition. The features extracted from the videos are fed into a k-NN
classifier, which classifies the level of congestion of a given scene as light, medium,
or heavy. The k-NN classifier was trained using annotated videos. However, as noted
by the authors themselves, this approach fails in distinguishing scenes with simi-
lar dynamics (for instance, confusion exists between empty road and stopped road
situations).
Another work based on optical flow analysis, but in the context of human action
recognition in video surveillance applications, is the one by Martinez et al. [ 20 ]. In this
work, dense optical flows using a method which captures multiscale information [ 19 ]
is computed. These optical flows are then used to build histogram-based descriptors.
An SVM is used to classify actions. Although not originally designed for traffic video
analysis, this work provides a general motion-based descriptor which might be used
for traffic scene analysis.
As an alternative to optical flow, Albiol and Mossi [ 1 ] have proposed to detect
corner points appearing in moving regions in order to estimate queue lengths. The
idea is that salient points such as corners normally result from the presence of cars and
that asphalt regions do not contain significant amounts of corners. Consequently, a
high number of corners is an indication of a densely occupied road. Motion detection
is used to obtain moving regions, and corners inside those regions are determined
via the Harris corner detector [ 9 ]. In their work, Albiol et al. also estimate the length
of the queues using a perspective estimation method. However, this method might
fail if the road is also highly textured.
13.3.2.2 Video Representation
The detection of moving objects is an important step for the video-based traffic
monitoring. Therefore, the first step was to separate moving object(s) from image
background using image segmentation according to features of the moving object(s).
There are three basic approaches for moving object detection which are background
Search WWH ::




Custom Search