01658302

Component

With the rapid development of urbanization, the safety of road intersections has been widely concerned. This paper presents an automated vision-based road user conflict detection system, which can provide more effective data for traffic safety diagnosis. The system can achieve the high-precision detection, classification and tracking of the road users by using the state-of-art deep convolution neural network and the MOT technology, and finally the potential traffic conflict events are identified by LSTM based trajectory prediction techniques and TTC indicator. The system was experimented on a typical intersection of Nanjing, where the conflict between vehicles and non-motor vehicles (PTW and bicycles) was detected and their safety conditions at the intersections were evaluated. The results showed that the method based on deep learning can better adapt to the conflict detection of complex intersection, and the safety of the intersection can be effectively analyzed by this method.

This paper was sponsored by TRB committee ABJ70 Standing Committee on Artificial Intelligence and Advanced Computing Applications. Alternate title: Traffic Conflict Detection of Vehicle and Nonmotorized Vehicle at Intersection Based on Deep Learning

18-00657

English

16p

2018

Transportation Research Board 97th Annual Meeting

Digital/other

Figures; Photos; References; Tables

Automatic incident detection; Bicycles; Computer vision; Human powered vehicles; Intersections; Machine learning; Neural networks; Nonmotorized transportation; Traffic conflicts; Traffic safety; Vehicle trajectories

Deep learning; Time to collision

Nanjing (China)

Highways; Pedestrians and Bicyclists; Safety and Human Factors; Security and Emergencies; Vehicles and Equipment

Transportation Research Board Annual Meeting 2018 Paper #18-00657

TRIS, TRB, ATRI

Jan 8 2018 10:10AM