Technical Services Group

EE Home Page


Student Resources

EE Students

EE Department Faculty

EE Faculty

EE Featured Projects

Featured Projects

EE Department Newsletters

EE Newsletter

EE Partnerships and Giving

Support EE

EE Photo Galleries

Support CENG

Automated Consensus-based Data Verification in the Caltrans Detector Testbed

Project Description

Image of car on highwayA practical need to assess the accuracy and attributes of each of the many types of roadway sensors and detectors motivated the California Department of Transportation to construct a traffic detector testbed on I-405 in Southern California.

With up to ten detectors of different types under concurrent test in each of six lanes, a means for automating the testing process became imperative, since traditional human verification methods were not practical. An automated data acquisition and verification system was built that utilizes a consensus of the results from all detectors under test, along with those of a reference image processing system, to create a preliminary ground truth record requiring manual verification for only a small percentage of ambiguous cases. Individual detector performance was then assessed by comparison with this verified dataset.

V2DVS (Video Vehicle Detector Verification System) - Test Results and Conclusions

  • Consensus-based data reduction greatly reduces the workload associated with ground truth generation, since it requires human verification only for detections that cannot be automatically correlated.
  • Automated data acquisition and reduction makes the processing of large numbers of vehicles over long test periods feasible.
  • In preliminary testing under various traffic and lighting conditions, approximately 97% of vehicle detections were properly classified as correct, false, or failure-to-detect by the automated process. Only 3% required human verification, and all of these cases were correctly identified.
  • Reference V2DVS image processing method is 99+% accurate under difficult ambient light conditions by use of shadow continuation as well as texture properties to discriminate actual vehicles from shadows.
  • Cases marked for human verification were most commonly related to ambiguous vehicle lane position or attributed to an excess of false detections by one or more of the detectors under test.
  • Automated data reduction accuracy is dependent upon the size of the admissible time/distance aperture, with more conservative settings tending to reject valid detections, while less conservative settings admitting incorrect matches.

Cal Poly filed a patent application on the V2DVS.

Contact Information

C. Arthur (Art) MacCarley, Ph.D., PE



Related Content