Citation

Qu L, Qu Z, Ren Z. Int. J. Veh. Safety 2023; 13(1): 64-89.

Copyright

(Copyright © 2023, Inderscience Publishers)

DOI

10.1504/IJVS.2023.137693

PMID

unavailable

Abstract

Estimating the real-time state of a vehicle is critical for driver assistance systems. In this study, we propose a framework that utilises deep learning (DL) and monocular cameras to obtain information about vehicles. We first estimate the velocity of the ego vehicle using optical flow. Then, the position of the visible vehicle is tracked and predicted using deepSORT and projection geometry. Finally, an extended Kalman filter (EKF) is used to combine the velocity, tracking information and relative position to estimate the trajectory of both vehicles in a common coordinate frame. We evaluated the accuracy of the visible vehicle's position using three public data sets. The results show that our proposed algorithm has a consistent Root Mean Square Error (RMSE) between 5.12 and 6.272. On two of the data sets, our algorithm outperforms other data-driven supervised DDSDL algorithms. To evaluate trajectory generation, we conducted both simulated experiments and real-world tests, demonstrating that our algorithm generates accurate vehicle trajectories.

Keywords: deep learning; relative position estimation; trajectory generation; non-linear optimisation.

Language: en