Computer Vision at Intersections: Explorations in Driver Assistance Systems and Data Reduction for Naturalistic Driving Studies
Authors
Philipsen, Mark Philip ; Jensen, Morten Bornø
Term
4. term
Publication year
2015
Submitted on
2015-06-16
Abstract
Denne kandidatafhandling, udført over næsten to semestre på UC San Diego inden for Vision, Grafik og Interaktive Systemer, undersøger automatisk detektion af trafiklys i billeder og video og udvikler desuden en køretøjsdetektor til vejkryds til brug i naturalistiske kørselsstudier. Vi gennemførte en omfattende gennemgang af forskning i trafiklysgenkendelse (TLR) i både akademia og industri og konstaterede mangel på udfordrende offentlige datasæt og standardiserede evalueringsmetoder (fælles test og mål). For at afhjælpe dette oprettede og offentliggjorde vi verdens største offentlige datasæt med cirka 110.000 annoterede (mærkede) trafiklys. Vi sammenlignede også læringsbaserede metoder med heuristiske, modelbaserede tilgange til trafiklysdetektion. Arbejdet har ført til indsendelse af én konferenceartikel til Intelligent Vehicles Symposium 2015, to konferenceartikler til Intelligent Transportation Systems Conference 2015 og en tidsskriftsartikel til Transactions on Intelligent Transportation Systems.
This master’s thesis, completed over nearly two semesters at UC San Diego within Vision, Graphics, and Interactive Systems, investigates automatic detection of traffic lights in images and video and also develops a vehicle detector for intersections used in naturalistic driving studies. We conducted a comprehensive review of traffic light recognition (TLR) research in academia and industry and found a lack of challenging public datasets and standardized ways to evaluate results (common tests and metrics). To address this, we created and released the world’s largest public dataset of about 110,000 annotated (labeled) traffic lights. We also compared machine learning–based methods with heuristic, model-based approaches for traffic light detection. The work led to the submission of one conference paper to the Intelligent Vehicles Symposium 2015, two conference papers to the Intelligent Transportation Systems Conference 2015, and a journal paper to Transactions on Intelligent Transportation Systems.
[This abstract was generated with the help of AI]
Documents