An Imaging Sensor-Aided Vision Navigation Approach that Uses a Geo-Referenced Image Database

Yan Li, Qingwu Hu, Meng Wu, Yang Gao
2016 Sensors  
In determining position and attitude, vision navigation via real-time image processing of data collected from imaging sensors is advanced without a high-performance global positioning system (GPS) and an inertial measurement unit (IMU). Vision navigation is widely used in indoor navigation, far space navigation, and multiple sensor-integrated mobile mapping. This paper proposes a novel vision navigation approach aided by imaging sensors and that uses a high-accuracy geo-referenced image
more » ... (GRID) for high-precision navigation of multiple sensor platforms in environments with poor GPS. First, the framework of GRID-aided vision navigation is developed with sequence images from land-based mobile mapping systems that integrate multiple sensors. Second, a highly efficient GRID storage management model is established based on the linear index of a road segment for fast image searches and retrieval. Third, a robust image matching algorithm is presented to search and match a real-time image with the GRID. Subsequently, the image matched with the real-time scene is considered to calculate the 3D navigation parameter of multiple sensor platforms. Experimental results show that the proposed approach retrieves images efficiently and has navigation accuracies of 1.2 m in a plane and 1.8 m in height under GPS loss in 5 min and within 1500 m. Sensors 2016, 16, 166 2 of 17 this approach does not depend on any signal or radiant sources [1, 6] . Nonetheless, this environmental condition becomes problematic when real-time and high-precision performance is required. Imaging sensor-based vision navigation estimates position and orientation information according to the geometrical relation from overlapping sequence images, which can improve the reliability of a navigation platform. DeSouza and Kak [7] investigated the developments in vision-related fields for mobile robot navigation over the past 20 years. Vision navigation is classified into three different categories: map-based, map building-based, and mapless navigation [7] [8] [9] . Mapless vision navigation utilizes imaging sensors and does not consider any prior description of the environment. Sequence images from the imaging sensors are used for motion analysis to determine the relative position and orientation information of a moving platform. The recently developed simultaneous localization and mapping (SLAM) technique integrates imaging sensors to extract the 3D navigation information with which 3D environment data are used for a map building-based vision navigation application. SLAM is employed in self-driving cars, unmanned aerial vehicles, autonomous underwater vehicles, planetary rovers, new domestic robots, and even within the human body [10] [11] [12] . At present, imaging sensors are combined with other laser scanning sensors to obtain robust positions and orientations for SLAM-based applications. Mapless vision navigation is a relative navigation technology that supports moving platforms to understand surroundings and explore a local environment, such as in indoor navigation and self-driving cars. This type of navigation stores information from the current environment, as well as its own relative position in the environment. Nonetheless, cases of absolute navigation with global 3D coordinates and orientation have been observed, such as outdoor navigation, the L-MMS for 3D surveying platforms, and space navigation [2, 13, 14] . These cases require global geo-referencing with highly accurate absolute 3D coordinates and may include GPS/IMU; however, harsh and poor environmental conditions limit the performance of GPS/IMU navigation. Thus, map-based vision navigation is implemented by providing a moving platform with a geo-referenced 3D model of the environment, such as a 3D map, landmark, and geo-referenced image database (GRID) [3, 15, 16] . Map-based vision navigation considers the map as a sensor to match the real-time imaging sensor and facilitate highly accurate absolute navigation in advanced driver assistance systems (ADAS). A collection of geo-referenced images serves as a map for real-time localization; this collection is normally treated as a 3D map because it contains not only a set of landmarks but also the corresponding 3D location information [17, 18] . Thus, imaging sensor-based vision navigation has two application scenarios as a map-based navigation approach: one is to improve the reliability of navigation under harsh environmental conditions in which GPS and IMU are ineffective, such as in ADAS. The other is to improve navigation accuracy where the vision navigation result is considered to be new input for Kalman filtering with GPS/IMU, such as in L-MMS. The key issues in image database-aided vision navigation are image searching and the matching methods with real-time images (RTIs) [19, 20] . Results of recent research show that the process of matching two images of the same scene under different scales, illumination, and view angles has been developed successfully for years; nonetheless, several theoretical and technical problems must be addressed [21, 22] . For example, two independent images share many similar features or pixels under certain circumstances, which may cause matching ambiguity given many candidate images. This condition is a serious problem for image database-based vision navigation. The speed of matching RTIs with the image database not only depends on feature computation but also on image search and retrieval. The organization model of the image database is combined with robust feature matching to realize matching with fast searching; these methods are the key techniques in GRID-based vision navigation. Our paper presents a novel imaging sensor-aided vision navigation approach that uses the highly accurate GRID. First, the framework of GRID-aided vision navigation is established with sequence images derived from land-based, multiple sensor-integrated mobile mapping systems. Second, a highly efficient GRID storage management model is developed based on the linear index of a road Sensors 2016, 16, 166 3 of 17 segment for fast image search and retrieval. Third, a robust image matching algorithm is presented to search for and match RTIs with GRID; the matched image is then matched with the real-time scene to calculate the 3D navigation parameter of the multiple-sensor platform. Experimental results show that the proposed approach retrieves images highly efficiently and has navigation accuracies of 1.2 m in a plane and 1.8 m in height under GPS loss in 5 min and within 1500 m.
doi:10.3390/s16020166 pmid:26828496 pmcid:PMC4801544 fatcat:e7wkx7yeofewvc4wct3usem7du