Surveillance cameras have been increasingly installed along roadways over last last
years to study traffic conditions. However, sight-degrading factors cause occlusion
which makes it extremely challenging to extract traffic parameters. Our approach
is designed to bypass complicated mathematical frame-to-frame vehicle segmentation
while accommodating the fact that vehicles in one lane can occlude another lane.
Our VBTMS does not track vehicles in each image but instead generates a graphical
representation of vehicle trajectories.
The algorithm consists of seven steps: (1) pre-processing including video stabilization
and (2) shadow removal, (3) pre-processing for image enhancement, (4)
camera calibration for image straightening, (5) time stacks for spatio-temporal map
generation, (6) trajectory segmentation and last (7) trajectory line approximation.
After extracting the trajectories, the occluders are excluded from neighboring lanes
using a trajectory matching algorithm. Therefore, corresponding features between
trajectories in lane 1 to lane 4 at the same point in time and space are checked for
equality.
Experimental results prove that our approach shows encouraging results, both
qualitatively in the ability to follow the video temporal slice evolution of traffic and
quantitatively when compared to speeds from single loop detectors.