The objective of this project was to investigate the use of existing video resources, such as traffic cameras, police cameras, red light cameras, and security cameras for the long-term, real-time collection of traffic statistics. An additional objective was to gather similar statistics for pedestrians and bicyclists. Throughout the course of the project, we investigated several methods for tracking vehicles under challenging conditions. The initial plan called for tracking based on optical flow. However, it was found that current optical flow—estimating algorithms are not well suited to low-quality video–hence, developing optical flow methods for low-quality video has been one aspect of this project. The method eventually used combines basic optical flow tracking with a learning detector for each tracked object– that is, the object is tracked both by its apparent movement and by its appearance should it temporarily disappear from or be obscured in the frame. We have produced a prototype software that allows the user to specify the vehicle trajectories of interest by drawing their shapes superimposed on a video frame. The software then tracks each vehicle as it travels through the frame, matches the vehicle’s movements to the most closely matching trajectory, and increases the vehicle count for that trajectory. In terms of pedestrian and bicycle counting, the system is capable of tracking these “objects” as well, though at present it is not capable of distinguishing between the three classes automatically. Continuing research by the principal investigator under a different grant will establish this capability as well. (Author/publisher)
Abstract