01590387

Component

The paper presents a novel computer vision-based traffic surveillance system capable of processing aerial imagery to track vehicles and their movements. The system uses a preprocessed 1-Hertz image sequence with a coverage of 25 square miles from an aerial camera array mounted on an airplane. The unique characteristics of the input data make this work challenging. Several heuristic and machine learning approaches are proposed and evaluated to track vehicles for the purpose of collecting traffic data. The system is capable of collecting speed, density, and volume data for uninterrupted flow corridors which is useful for “big data” monitoring of traffic parameters over an entire 25 square mile area with a single sensor. The deep learning and SURF-based approach is able to achieve over 93% accuracy throughout 50 seconds on density estimates when compared with manually collected ground truth. It has 100% accuracy when level of service (LOS) was measured for the uninterrupted facilities tested. These evaluations were conducted for facilities of different levels of congestion as indicated by the different levels of service. With further research, improved preprocessing, and a higher frame rate, the accuracy of tracking vehicles can be improved which will allow for other potential applications such as identification of erratic drivers and origin-destination studies.

This paper was sponsored by TRB committee ABJ50 Standing Committee on Information Systems and Technology.

01584066

16-6783

English

17p

2016

Transportation Research Board 95th Annual Meeting

Digital/other

Figures; References; Tables

Aerial photography; Artificial intelligence; Automatic vehicle detection and identification systems; Computer vision; Data collection; Data mining; Traffic surveillance

Data and Information Technology; Highways; I72: Traffic and Transport Planning

Transportation Research Board Annual Meeting 2016 Paper #16-6783

TRIS, TRB, ATRI

Jan 12 2016 6:57PM