Visual Odometry with Dynamic Object Detection by Complementary Integration of Optical Flows and Pattern Recognition

Kojiro Takeyama, Takashi Machida, Yoshiko Kojima, Nobuaki Kubo

Abstract:	In this paper, a method of vehicle trajectory estimation through accurate dynamic object detection is proposed. Robust and accurate vehicle position information is essential in driver assistant systems. Further, vehicle trajectory estimation is important for vehicle positioning in areas where GPS is not available, such as urban canyons. Visual odometry (VO) using mono-camera is a potentially low cost and high performance trajectory estimation method. However, in urban environments its accuracy is degraded because of dynamic objects such as vehicles. The proposed method overcomes this challenge using a two-step solution. In the first step, dynamic object detection based on optical flow is performed using IMU as a reference motion of the vehicles. This enables avoidance of large errors and contributes to ensuring the integrity of the refinement process in the second step. The second step, the key point of this study, is complementary integration of dynamic object detection based on optical flow and pattern recognition. This enables accurate dynamic object detection even when both steps have some amount of detection error. The results of evaluations conducted in a city indicate that the proposed method reduces trajectory error by as much as 71% compared with conventional methods.
Published in:	Proceedings of the 30th International Technical Meeting of the Satellite Division of The Institute of Navigation (ION GNSS+ 2017) September 25 - 29, 2017 Oregon Convention Center Portland, Oregon
Pages:	3301 - 3310
Cite this article:	Takeyama, Kojiro, Machida, Takashi, Kojima, Yoshiko, Kubo, Nobuaki, "Visual Odometry with Dynamic Object Detection by Complementary Integration of Optical Flows and Pattern Recognition," Proceedings of the 30th International Technical Meeting of the Satellite Division of The Institute of Navigation (ION GNSS+ 2017), Portland, Oregon, September 2017, pp. 3301-3310. https://doi.org/10.33012/2017.15316
Full Paper:	ION Members/Non-Members: 1 Download Credit Sign In