Airborne Video Registration and Traffic Flow Parameter Estimation in a Multimedia GIS

Investigators

Robert Schowengerdt, Electrical and Computer Engineering Department, U. Arizona

Pitu Mirchandani, Systems and Industrial Engineering Department, U. Arizona

Mark Hickman, Department of Civil Engineering and Engineering Mechanics, U. Arizona

Student

Anand C. Shastry, Electrical and Computer Engineering Department, U. Arizona

Airborne video data acquisition for traffic parameter estimation has been explored as an alternative to conventional data collection methods. We have developed a computer vision system for spatially registering the video and estimating traffic flow parameters from the video. Our registration algorithm makes use of the KLT feature tracker to automatically extract feature point correspondences and completely eliminates human intervention, making it possible to process hundreds of frames in succession. The registered video has thus eliminated unwanted platform motion, without making use of any telemetry or elevation data. We have implemented a novel window sizing scheme for the KLT feature tracker that avoids finding control points on the moving vehicles during registration, making the registration process very robust. Vehicle detection is performed by simple frame differencing and motion detection in the registered sequence. A different window size corresponding to the dimensions of vehicles is used to track the detected vehicle locations through the sequence, allowing us to automatically compute the velocity of vehicles, instantaneous velocity and the vehicle trajectories. Processing the video frames using the algorithm produces a file with time varying vehicle coordinate locations that is used to plot a vehicle space-time diagram and can be used for further analysis and parameter estimation.

The registered video is also a good visualization tool, as it completely removes platform motion, leaving only the vehicles moving. We have created an extremely useful tool for visualization, by extending the capabilities of a traditional GIS to include multimedia data accessible from a user-friendly interface. To facilitate comprehensive data handling, the registered video and other imagery are located in the multimedia GIS. To help focus attention of traffic analysts on the roadways, there is a digitized roadway layer in the GIS that masks out the non roadway segments. This has been placed over a black and white orthophoto, to help easy identification of the city locations.

The registration algorithm was successfully tested for several video clips taken from a hovering, but shaky helicopter, as well as video of the helicopter following vehicles on arterial roads. By modifying the registration transformation, it was possible to correct for platform attitude changes without removing the constant velocity motion of the helicopter. The jitter was on the order of a few pixels over several hundred frames. The algorithm was able to correctly detect about 65% of the vehicles. The modified feature tracker tracked about 90% of the vehicles accurately. Velocity estimation was within 10% of the manual measurements, usually much closer. Vehicle movement trajectories and instantaneous velocity were also obtained, making it useful for transportation analysts. The computation time on a Sunblade 2000 (UltraSPARC3 900MHz) is of the order of a few seconds for processing each frame. Thus, in less than an hour, a twenty second video clip can be registered, the vehicles detected, tracked in time and the statistics obtained – all completely automatically. (Excerpt from Anand C. Shastry, “Airborne video registration and traffic flow parameter estimation in a multimedia GIS”, M.S. thesis, 2002, Department of Electrical and Computer Engineering, University of Arizona.)

This research was sponsored by the US Department of Transportation Other Transaction Agreement DTRS56-00-T0004 on Remote Sensing of Transportation Flows.

Here are a few helicopter videos available for download. These are compressed videos in avi format. The videos show results of the registration and vehicle tracking and are in pairs showing the original and corresponding processed videos.