This project will expand on previous work that created a new algorithm to estimate speed using a sequence of video images from an uncalibrated camera. The original algorithm used frame differencing to isolate moving edges and track vehicles between frames. It combined a known vehicle length distribution with image information to estimate speed. The expanded research will investiage the automation of the algorithm to create a new speed sensor from the existing CCTV cameras deployed by WSDOT.
The large numer of roadside cameras used by departments of transportation are typically not installed in a manner that allows them to be easily calibrated. Also, they are typically used by operators who can tilt, pan, and zoom with a joystick to change the camera viewing area. The combination of movable cameras and a lack of calibration makes estimating speed for uncalibrated cameras a challenge.
The ablity to automatically select the subset of vehicles to be tracked is the primary difficulty in automating the existing algorithm. Initially, probe vehicles will be used to create the image sequences for verifying correct tracking and speed estimates. Once the algorithm has been validated against probe vehicles, inductance loop sensors will be used to test the accuracy of the uncalibrated camera speed sensor relative to operating traffic management sensors. Standard statistical tests will be used to compare the speed estimates from these sources and to quantify the errors and biases inherent in using this type of camera speed sensor for operational traffic conditions.
This research will allow quantitative speed estimates from the existing camera pool and will eliminate the need for additional, specially calibrated cameras and associated infrastructure. If successful, it will create prototype software and algorithms capable of producing speed estimates from cameras that output 3 image frames per second.