Vehicle Detection and Identification Algorithm Tested in Columbia, South Carolina Found to Increase Video Processing Efficiency at Reasonable Costs.
Columbia, SC, United States
Summary Information
The collection of traffic volume data is vital to transportation agencies for various purposes. A range of techniques exist for counting traffic volumes including radar detectors, inductive loops, infrared axle detectors, side-fire radar and video-based systems. The primary objective of this study was to develop image processing algorithms capable of automatically extracting vehicle counts and classifications, as well as counts of motorcycles, bicycles, and pedestrians from videos. As part of the study, a survey of 19 Departments of Transportation (DOT) was conducted to obtain information regarding the use of video-based systems for traffic counting and classification. Only three out of the 11 DOTs that use video-based systems had the vehicle classification task performed in-house; the rest outsourced this work to a vendor. The second objective of the study was to incorporate the developed algorithms into a stand-alone application with an easy-to-use interface to enable the South Carolina DOT (SCDOT) staff to process traffic videos in house.
METHODOLOGY
This study developed algorithms to classify vehicles in traffic videos, using background subtraction and foreground detection methods along with a Convolutional Neural Network (CNN) model. The developed algorithms and CNN model were integrated into a user-friendly application called DECAF, allowing users to easily specify video files, region, and time interval for analysis. DECAF was tested with videos from five different sites in Columbia, SC, achieving over 95 percent accuracy in both detection and classification for roads with up to four lanes.
FINDINGS
Using the in-house equipment and software, the video processing cost per 48-hour count was approximately $500.
Automatic Extraction of Vehicle, Bicycle, and Pedestrian Traffic from Video Data
See summary