Citation

Liu X, Wang H, Dong C. Int. J. Intell. Transp. Syst. Res. 2021; 19(2): 312-323.

Copyright

(Copyright © 2021, Holtzbrinck Springer Nature Publishing Group)

DOI

10.1007/s13177-020-00247-w

PMID

unavailable

Abstract

High computational cost and low tracking stability make it still a challenging task to acquire nonmotorized traffic parameters at intersections via vision-based method. In order to address the above issues, our study improves a cooperative tracking and classification method, and proposes a vision-based data collection system to monitor nonmotorized traffic at intersections. The system utilizes the combination of two tracking algorithms, Kernelized Correlation Filter and Kalman filter, to ensure the continuous tracking. Based on multivariate feature, K-means clustering and Support Vector Machine are implemented to classify nonmotorized traffic according to the motion and appearance feature respectively. As a result, the proposed system can acquire trajectories of pedestrians and cyclists and extract traffic parameters, including flow and velocity. Our method performs well in both efficiency and accuracy by fusing simple but effective algorithms and is robust in the complex scenario especially at large-scale intersections with limited training samples. The experimental results show that it can extract more trajectories with low computational lost. Moreover, the error of flow and velocity result is controlled within acceptable limits, which directly proves it feasible to collect field data in project applications.

Language: en