Intermodal Mobility Assistance for Megacities - Smart Information Systems: Computational Intelligence for Real-Life Applications - page 352

precisely detect vehicles. Tracking is achieved via Kalman filtering. Their solution
is used to count vehicles and detect jams. However, according to their results, traf-
fic jam detection is limited to identifying stationary vehicles, and the videos used
for testing only show fluid traffic. Therefore, it is not clear if the method in [ 14 ]is
scalable to congestion analysis in challenging situations.
In a similar manner, Wu et al. [ 27 ] use motion information for detecting vehicles.
In addition to motion, they use edges located in those regions showing motion.
An analysis of this edge map using heuristic rules results in the determination of
individual vehicles. Vehicles are subsequently tracked. The authors have applied
their algorithm to the determination of congestion levels. However, this approach
shows serious limitations, in particular the detection of vehicles, which is based on
heuristics. For instance, the authors make assumptions about the size of vehicles.
Consequently, this method is not easily scalable, as it needs to be calibrated for each
camera or scene setup.
A more advanced solution to vehicle detection and classification was proposed by
Buch et al. who developed a method which makes use of motion information [ 3 ]. In
their work, Buch et al. use the Stauffer-Grimson Gaussian Mixture Model to extract
the moving objects (blobs) appearing in the video. Individual blobs are verified
against 3-D wire frame models in order to classify vehicles.
However, although object detection and tracking have nowadays reached an
unprecedented level of performance, detection and tracking are, generally speaking,
not yet fully appropriate and operational in the context of traffic video analysis where
videos are often lower in resolution, and where vehicles often appear occluded due to
viewpoints. For instance, segmentation of individual vehicles might be problematic
or tracking might fail because of such occlusions. Therefore, an increasing number
of researchers of the traffic video analysis community have looked for alternatives
represented by the methods of the second category.
One example is the solution proposed by Lee and Bovik [ 13 ], who adopted an
approach which does not require individual detection and tracking of vehicles. Their
solution is based on the global analysis of optical flow of traffic videos, followed by
a statistical analysis of flow regions to extract meaningful information. Optical flow
is estimated via a robust gradient-based solution [ 2 ] which allows a representation of
different traffic lanes appearing in a traffic video. The subsequent statistical analysis
consists in a histogram analysis of flow vectors. However, the approach was only
illustrated in the context of flow anomaly detection. For instance, no determination
of congestion levels was demonstrated.
Sobral et al. [ 24 ] have proposed a method based on vehicle crowd density esti-
mation and tracking, which does not rely on individual vehicle detection or tracking.
First, the extraction of moving blobs by background subtraction is used to deter-
mine crowd density. Second, crowd tracking by the Kanade-Lucas-Tomasi (KLT)
tracker [ 23 ] is performed to estimate the speed of vehicle crowds. The analysis is
limited to a region of interest (ROI) of size 190 by 140 pixels. The resulting crowd
density and speed are concatenated into a feature vector. Classification of feature
vectors is performed via various classifiers including k-nearest neighbors (k-NN),
support vector machines (SVM) and neural networks (NN), and returns the level of