Automatic Extraction of Vehicle, Bicycle, and Pedestrian Traffic From Video Data

SPR No. 742This project investigated the use of traffic cameras to count and classify vehicles. The intent is to provide an alternative approach to pneumatic tubes for collecting traffic data at high volume locations and to eliminate safety risks to SCDOT personnel and contractors. The objective is to develop algorithms to post-process the 48-hour videos to determine the number of vehicles in each one of four categories: motorcycles, passenger cars and light trucks, buses/campers/tow trucks, and small to large trucks. To this end, background subtraction and foreground detection algorithms were implemented to detect moving vehicles, and a Convolutional Neural Network (CNN) model was developed to classify vehicles using thermal images obtained from a custom-built thermal camera and solar-powered trailer. Additionally, to overcome false detection of vehicles due to either camera motion or erratic light reflection from the pavement surface, an algorithm was developed to keep track of each vehicle\u2019s trajectory and the vehicle trajectories were used to determine the presence of an actual vehicle. The developed algorithms and CNN model were incorporated into a Windows-based application, named DECAF (detection and classification by functional class) to enable users to easily specify the folder that contains the video files to be processed, specify the region for which traffic should be analyzed, specify the time interval for which the data should be aggregated, and view the detection and classification results in two report formats: 1) a spreadsheet with vehicle-by-vehicle information, and 2) a PDF summary report with totals aggregated for the user-specified interval. DECAF was tested using videos collected from five different sites in Columbia, SC, and the overall detection and classification accuracy for the hours evaluated was found to be 95% or higher