Real Time Classification of Vehicle Types and Modes using Image Analysis and Data Fusion

2019PDFTech ReportHuynh, Nathan N.Mullen, Robert LMejia, YohannaUniversity of South Carolina. Dept. of Civil & Environmental EngineeringCenter for Connected Multimodal Mobility, Clemson UniversityCenter for Connected Multimodal Mobility, Clemson UniversityUnited StatesData fusionImage analysisMethodologyNeural networksTraffic countsVehicle classificationVehicle detectorsComputer visionCascadeClassifiersTransportation modesLocal Binary PatternsHaar-like features69A3551747117Final Report January-December 2018Several methodologies to count and classify vehicles from video recordings were explored in this research to provide transportation agencies with a safer alternative to collect these data which are needed for traffic engineering and planning studies. In this study, three approaches for detecting and classifying vehicles were investigated: Haar-like features, Local Binary Patterns (LBP), and Convolutional Neural Networks (CNN). Recordings from the South Carolina Department of Transportation (SCDOT) were used for the study. By means of image subtraction and contouring, specific areas of the video stream were extracted as independent images. The resulting database was used for training and testing the Cascade classifiers and the CNN. Results indicate that using a data set consisting of 5,500 car and 2,800 truck images, the CNN achieved higher accuracy (97%) in vehicle classification compared to LBP and Haar cascade.104