Truck Taxonomy and Classification using Video and Weigh-In-Motion (WIM) Technology
The primary objective of the present work was to develop video processing and machine learning methods to automatically detect and classify trucks traveling on Florida highways. The assembly of an automated system to detect, classify, and recognize various truck types from video and to derive their attributes is presented in this report. These extracted attributes were then used to determine commodity types, which can be used downstream for tracking commodity movements. To accomplish this, a set of high resolution videos (made available by the Florida Department of Transportation, FDOT) using freeway roadside passive cameras was utilized to create benchmark datasets. These videos were captured at different times of day, mainly at two freeway locations, and on various days during the past two years. The set of images derived from the videos was processed by the authors' developed system to train and evaluate different approaches. The approaches drew upon recent work in deep convolutional neural networks for object detection and classification, semantic segmentation, and feature extraction, as well as drawing from traditional methods such as decision trees and geometric features (like edges and corners). The authors developed deep learning algorithms that leveraged transfer learning to determine whether an image frame has a truck and, if the answer is affirmative, localize the area from the image frame where the truck is most likely to be present. In particular, (1) they developed deep learning algorithms for detecting the location of a truck in a video frame followed by determining whether the image corresponds to a truck or a non-truck, (2) they developed a hybrid truck classification approach that integrates deep learning models and geometric truck features for classifying trucks into one of the nine FHWA classes (FHWA classes 5 through 13), (3) they developed algorithms for recognizing and classifying various truck attributes such as tractor type, trailer type, and refrigeration units that are useful in commodity prediction, and (4) they developed techniques for extracting vendor information corresponding to a truck, using logo and text detection.
- Record URL:
- Summary URL:
- Summary URL:
- Record URL:
-
Corporate Authors:
University of Florida, Gainesville
Department of Computer & Information Science & Engineering
Gainesville, FL United States 32611Florida Department of Transportation
605 Suwannee Street
Tallahassee, FL United States 32399-0450 -
Authors:
- He, Pan
- Wu, Aotian
- Rangarajan, Anand
- Ranka, Sanjay
- Publication Date: 2019-5
Language
- English
Media Info
- Media Type: Digital/other
- Edition: Final Report
- Features: Appendices; Figures; Photos; References; Tables;
- Pagination: 76p
Subject/Index Terms
- TRT Terms: Classification; Machine learning; Neural networks; Taxonomy; Trailers; Trucks; Vehicle classification; Weigh in motion
- Geographic Terms: Florida
- Subject Areas: Data and Information Technology; Highways; Motor Carriers; Planning and Forecasting; Vehicles and Equipment;
Filing Info
- Accession Number: 01718168
- Record Type: Publication
- Contract Numbers: BDV31-977-81
- Files: NTL, TRIS, ATRI, STATEDOT
- Created Date: Sep 25 2019 5:17PM