We present an algorithm to estimate mean vehicle speed from roadside cameras operated by a traffic management agency. These roadside cameras are neither calibrated nor are there calibration marks available in the camera views. However, estimating camera calibration coefficients is the most important step to extracting quantitative information about the 3D world from a 2D image. It is in this framework that we present an algorithm that: (1) performs a simplified dynamic calibration and (2) estimates mean vehicle speed. Many algorithms depend on point correspondences between the earth coordinates and the image coordinates as well as targets of known shape to obtain accurate results. However, in the work presented, we desire to estimate the mean of a distribution of vehicle speeds and will demonstrate that a simplified form of calibration is adequate for making an accurate mean speed estimate. We perform dynamic camera calibration using training sets of 10-second video sequences. Our proposed method detects moving vehicles in a set of consecutive frames. This information, together with mean vehicle dimension estimates, is used to create scaling factors that are then used to transform between motion in the image and motion in the earth coordinate system. Our proposed algorithm has a camera model with a reduced number of camera calibration parameters. We validate our algorithm with simulated data and real-world traffic scenes.