As traffic cameras become prevalent, and a considerable
amount of traffic videos are stored for various purposes, new possibilities
and challenges open in the automatic analysis of traffic scenes. Advances
in deep learning also enable new ways to characterize traffic in such
videos automatically. This work is motivated by the need to understand
traffic flow without human supervision, especially the localization of road
intersections in scenes from traffic cameras. For this purpose, a method
is proposed that uses a deep learning neural network for vehicle detection, an object tracker to recover vehicle trajectories from the detections,
and unsupervised machine learning techniques to detect potential incoming and outgoing traffic flows from the vehicle trajectories in the video
sequences. A wide range of real and synthetic videos have been used to
test the goodness of the proposal with satisfactory results, from traffic
cameras at different heights and angles, different traffic patterns, and
various weather conditions.
