Integration of vision and topological self-localization for intelligent vehicles

Self-localization is a crucial task for intelligent vehicles. Conventional localization methods usually suffer from different limitations, such as low accuracy and blind areas for Global Positioning System (GPS), high cost for Inertial Navigation System (INS), and low robustness for vision Simultaneously Localization and Mapping (vSLAM). To overcome these problems, this study proposes a low-cost yet accurate localization method for intelligent vehicles, which only needs a monocular camera and a GPS receiver. First, the proposed method offers multiple feature spaces which are designed from GPS data, localization prediction model, image holistic features, and image local features. Each feature space, from which one candidate set is derived, can make qualitative localization achieved independently. Afterwards, we propose a novel method called K-Nearest Neighbors from Multiple Feature Spaces (KNN-MFS) to fuse these candidate sets. The closest node to the current vehicle position is drawn from the visual map to achieve image-level localization. Finally, the vehicle pose in the visual map, computed by metric localization, can further enhance the localization accuracy. The advantage is that when GPS signals are unavailable at times, the method can still achieve short-range localization by using other feature spaces. The proposed method has been validated with the actual data sets and public data sets. The actual data sets are collected along an industrial park and a rural urban fringe in Wuhan City, China, covering different times and weather conditions, and the total lengths of these routes have to be more than 8 km. The public data sets are Karlsruhe Institute of Technology and Toyota Technology Institute (KITTI). Experimental results show that the proposed method can adapt to different times and weather conditions with good robustness in varying environments, and the localization errors are less than 35 cm in all the tests in average. Experimental results on the same routes without GPS data are also reported, which demonstrate that the proposed method can achieve comparable localization accuracy.