Computer Engineering and Applications ›› 2022, Vol. 58 ›› Issue (23): 197-204.DOI: 10.3778/j.issn.1002-8331.2105-0313

• Pattern Recognition and Artificial Intelligence • Previous Articles     Next Articles

Research on Dynamic Feature Visual Odometry in Road Environment

YANG Binchao, XU Xinying, CHENG Lan, FENG Zhou   

  1. 1.College of Electrical and Power Engineering, Taiyuan University of Technology, Taiyuan 030024, China
    2.Shanxi Key Laboratory of Advanced Control and Equipment Intelligence,Taiyuan 030024, China
  • Online:2022-12-01 Published:2022-12-01

道路环境下动态特征视觉里程计研究

杨斌超,续欣莹,程兰,冯洲   

  1. 1.太原理工大学 电气与动力工程学院,太原 030024
    2.先进控制与装备智能化山西省重点实验室,太原 030024

Abstract: To deal with the navigation and positioning of vehicles in the road environment, a monocular visual odometry based on the dynamic semantic features of the road environment is proposed. A self-supervised convolutional neural network is designed to model the monocular continuous image, no longer rely on artificially designed feature and directly predict the depth map and the pose vector. To deal with the destruction of photometric consistency by dynamic objects in the road environment, it is proposed to use semantic prior to improve the accuracy of visual odometry. Two fully connected layers are designed to estimate the rotation and translation vectors respectively. The experimental results show that the method is comparable to that of the traditional method, and it has superior performance in the road environment.

Key words: visual odometry, semantic segmentation, self-supervision, convolutional neural network, monocular image

摘要: 针对道路环境下移动车辆导航和定位的问题,提出了一种基于道路环境动态语义特征的单目视觉里程计。设计了一个自监督的卷积神经网络,对单目连续图像建模,直接预测深度图和位姿向量,不再依赖人工设计的特征点。针对道路环境下动态物体破坏光度一致性的问题,提出利用语义先验信息提高视觉里程计精度。设计两个全连接层分别估计旋转和平移向量。实验结果表明,该方法得到了与传统视觉里程计比肩的精度,并且在道路环境下具有优越的性能。

关键词: 视觉里程计, 语义分割, 自监督, 卷积神经网络, 单目图像