Algorithms for estimating mean vehicle speed using uncalibrated traffic management cameras
Todd N. Schoepflin, Yongmin Kim, Daniel J. Dailey · 2003
In this thesis we contribute a variety of algorithms for a space mean speed sensor that processes digital video of a road captured by an uncalibrated camera. We introduce a simplified model of the camera and the road scene and derive three camera calibration methods. For each of these methods, we perform a sensitivity analysis and determine the range of operation for the camera, assuming realistic levels of input noise. The three calibration methods rely on estimates of multiple parameters from the set of images, e.g., the road boundary lines and their vanishing point. We describe how to extract features from the images to estimate these parameters and their confidence intervals. On occasion, a given scene is impossible to analyze due to obstructions between the camera and the road (e.g., raindrops on the camera lens). We describe an algorithm to determine if the image features are unreliable. Assuming the camera has been calibrated and the vehicle lanes have been identified, we present a method to track a group of vehicles in a lane and estimate the mean speed using a cross-correlation technique. The algorithm is appropriate for building a sensor with fine time resolution (i.e., 200 ms); 20-second averages are shown to be equivalent to data from two different inductance loops. The results for several test cases show that the speed estimation method performs well under a variety of challenging weather, lighting, and traffic conditions. Finally, we analyze the computational requirements of the set of algorithms and conclude that a mediaprocessor-based system should achieve real-time operation on 320 x 240 images sampled at 5 Hz.