Robust Automatic Detection of Traffic Activity

The advent of autonomous vehicles has driven the need for high-performance traffic environment perception systems. In this context, streaming perception, which involves detecting and tracking objects in a video stream simultaneously, is a fundamental technique that significantly impacts autonomous driving decision-making. Notably, the fast-changing scale of traffic objects due to vehicle motion can lead to conflicts in the receptive field when detecting both large and small objects. Moreover, real-time perception is an ill-posed problem that heavily depends on motion consistency context and historical data. Consequently, two major challenges in real-time perception are: (1) adaptively handling rapidly changing object scales, and (2) accurately and efficiently learning long-term motion consistency. In this report, the authors have developed two systems for enhancing traffic safety. The first system focuses on road activity detection, which identifies the activities of vehicles. The authors discuss the first system, Argus++, in Chapter 2. Further, the authors integrated the models in their video analysis framework Argus++ to enable the real-time processing of traffic footage, including vehicle tracking. The authors introduce it in Chapter 4. Immediate notification could be provided on traffic density and speed estimation, and traffic incident detection. The second system focuses on streaming perception, which enhances the safety of autonomous driving. For this system, the authors introduce the models and algorithms in Chapter 3. Further, in Chapter 5, the authors discuss how to train a powerful model for their systems, especially training vision-language transformers from captions. The authors showed that this model could enhance the vehicle detection task.

• Record URL:
  • https://ppms.cit.cmu.edu/media/project_files/403_-_Final_Report.pdf
• 
• Summary URL:
  • https://ppms.cit.cmu.edu/projects/detail/403
• Record URL:
  • https://rosap.ntl.bts.gov/view/dot/68085
• Supplemental Notes:
  • This document was sponsored by the U.S. Department of Transportation, University Transportation Centers Program.
• Corporate Authors:

  National University Transportation Center for Improving Mobility (Mobility21)

  Carnegie Mellon University
  Pittsburgh, PA  United States  15213

  Office of the Assistant Secretary for Research and Technology

  University Transportation Centers Program
  Department of Transportation
  Washington, DC  United States  20590
• Authors:
  • Hauptmann, Alexander
  • 0000-0003-2123-0684
  • Yu, Lijun
  • 0000-0003-0645-1657
  • Liu, Wenhue
  • 0000-0003-4679-2958
  • Qian, Yijun
  • 0009-0003-9440-9744
  • Cheng, Zhiqi
  • 0000-0002-1720-2085
  • Gui, Liangke
  • 0009-0009-0338-8948
• Publication Date: 2023-6-30

Language

• English

Media Info

• Media Type: Digital/other
• Edition: Final Research Report
• Features: Appendices; Figures; Glossary; Photos; References; Tables;
• Pagination: 73p

Subject/Index Terms

• TRT Terms: Automatic data collection systems; Automatic tracking; Autonomous vehicles; Information processing; Machine vision; Vehicle detectors; Video
• Subject Areas: Data and Information Technology; Highways; Operations and Traffic Management; Vehicles and Equipment;

Filing Info

• Accession Number: 01887420
• Record Type: Publication
• Contract Numbers: 69A3551747111
• Files: UTC, NTL, TRIS, ATRI, USDOT
• Created Date: Jul 17 2023 9:13AM