Previous Abstract Return to Session D4

Session D4: Ground Vehicle Navigation

Robust Vehicle Localization with Collaborative Mapping of Road-Infrastructure Objects
Fabian de Ponte Müller and Omar García Crespillo, German Aerospace Center (DLR), Germany
Location: Spyglass

Introduction & Motivation
Future automated and autonomous vehicles will require highly precise and robust localization at any time and in any condition. Research work in the past decades has demonstrated that the use of several independent localization technologies is the key to success. Global Navigation Satellite Systems (GNSS) represent the corner-stone to pin the local environment to a global coordinate system. However, the quality of GNSS-based positioning in a land-based environment drops dramatically under unfavorable satellite geometries and high number of blocked satellites, in the presence of strong multipath components or due to the reception of non-line-of-sight (NLOS) signals. Therefore, solutions including on-board inertial or kinematic sensors, as for instance accelerometers, gyroscopes, wheel-tick sensors or steering angle sensors, have been favored to support GNSS in these kinds of environments. In recent years, many research groups have worked on vehicle localization algorithms using on-board ranging sensors in addition to baseline GNSS/INS integration [1, 2]. Laser-scanners, radar sensors and cameras are also used to effectively contain the integrative effect of errors in inertial-based localization. In many cases, these technologies make use of highly-detailed, ultra-precise and heavy maps that are acquired by a separate vehicle in advance [3, 4]. These maps have to be downloaded to the vehicles to be used for the localization process. This requires some sort of high data rate communication infrastructure, whose coverage might not be guaranteed globally. Moreover, since these maps are built in advance, they cannot respond to changes in the road environment that happen due to parked cars, temporary construction sites or snow piles. This limits the accuracy of the map and increases the risk of miss-association.
In our previous paper [5], we proposed a concept using discrete Road Infrastructure Objects (RIO) detected inside a GNSS-denied area by using an on-board ranging sensor, i.e. laser or radar. Within a simulation environment we showed that successive vehicles traversing this area and sharing limited information about detected RIOs can build a cooperative map that potentially enables sub-meter level position accuracy. The main benefit of this approach relies in its simplicity. Instead of requiring high definition 3D maps, a set of discrete RIO positions is sufficient to enable continuous vehicle localization. It also makes it possible to share the map directly with standard vehicle-to-vehicle (V2V) ad-hoc communication, as for instance Dedicated Short Range Communication or ITS-G5. After traversing a GNSS-impaired area, a vehicle will share the map with oncoming vehicles that will carry it and disseminate it to vehicles entering the GNSS-impaired area.
In this paper, we move from the simulation to the real-world, proving that the cooperative infrastructure-based localization of vehicles is able to enhance and maintain a bounded position accuracy in GNSS-denied and other challenging environments.
Cooperative Infrastructure Localization & GNSS Fault Detection Mechanism
The proposed localization algorithm is based on an Extended Kalman Filter where we integrate GNSS in a tightly-coupled fashion with laser and radar measurements and inertial navigation system (INS). The filter estimates therefore the error states of the INS along with the GNSS receiver clock. A limited number of discrete road infrastructure objects are also considered as states. The cooperative localization is achieved when the vehicles share the discrete map of road objects along with their associated uncertainties with other vehicles, so that vehicles in the vicinity can update and enhance their own maps. This collaborative process aims at improving the accuracy of the map over time and at responding to changing environmental road objects, as for instance parked cars or road construction sites.
The Kalman filter is designed under the assumption that the sensor errors are well modeled by certain Gaussian distributions. In the case of high multipath, the pseudoranges for instance will present different error distributions or would belong to a different one. In the case of wrong data association of road objects, the considered map measurement would not belong either to the modeled distributions. In order to make the filter more resilient against these types of faults, we also include on one side a local fault detection mechanism for the GNSS measurements and we design the data association of road objects in order to minimize the risk of wrong association.
Field Measurements
For the real-world experiments a test vehicle has been equipped with various sensors. For detecting the road infrastructure objects, a radar sensor and a laser scanner were installed on the front side of the vehicle. Two GNSS receivers, a medium-cost and a high-end geodetic grade receiver were used for the reference system and for the cooperative infrastructure-based positioning algorithm, respectively. Two inertial systems, a Microelectromechanical (MEMS)-based Inertial Measurement Unit (IMU) and a fiber-optic gyroscope have been attached to the vehicle chassis.
The experiments were performed in a commercial area in the outskirts of the city of Munich. In total 15 repetitions of a predefined 3 km drive in this environment has been recorded. The purpose is to emulate a drive of 15 independent vehicles passing along the same section of road. Light poles and trees located on both sides of the road will serve as road objects that are used by the localization algorithm to construct a map.
Expected Results
The main goal of this work is to determine the benefit of the cooperative sharing of local map information. In a first step, the ground truth of the road infrastructure map is constructed by computing the position and orientation of the vehicle by post-processing the data from the high-end geodetic-grade GNSS receiver and the fiber-optic gyroscope and computing the position of stand-alone road objects, as for instance light poles and trees, in a global coordinate frame. We then evaluate the iterative creation of the road infrastructure map, by on-line positioning each of the 15 vehicles and updating their map information with the received map from previous vehicles. This is done for both the laser scanner and the radar sensor independently. Special emphasis will be put on the analysis of miss-detection and miss-association of road objects, which represents a high thread to the resilience and convergence of map-based localization algorithms. Further on, the number of detected road objects incorporated into the map is limited by the packet-size and the channel capacity given by the corresponding V2V communication technology. This impact on the localization algorithm will be part of the analysis as well.
We first analyze the achieved performance in a GNSS nominal situation, that is, with favorable sky visibility and where there is no presence of strong multipath components or NLOS signals. We then test our algorithm in a GNSS-challenged scenario by deactivating single satellites, emulating multipath propagation errors or simulating a complete GNSS blockage. The main outcome is, first, the achieved accuracy improvement due to the high multisensor redundancy. Of particular interest is how the laser scanner and radar measurements help containing the growing inertial errors when only few satellites are visible. Second, we analyze also the improvement in terms of GNSS signal fault detection. Thanks not only to the inertial redundancy but also to the on-board ranging measurements, the detection capability of multipath corrupted pseudoranges and NLOS is increased.
Conclusions
With this paper we aim to answer the following three questions:
• Is it possible to collaboratively learn the location of surrounding road-infrastructure objects detected with on-board ranging sensors and use this for accurate and robust ego-location of vehicles? Which of the two sensors, the laser scanner or the radar, offers the best performance in terms of precision, accuracy and robustness?
• How should road features be selected in order to minimize miss-associations while maintaining an acceptable degree of position uncertainty? What is a practical number of road features that can still be exchanged over ad-hoc V2V communication?
• What is the improvement of GNSS fault detection and exclusion algorithms when using road-infrastructure measurements from on-board ranging sensors like laser scanners or radar sensors?
[1] A. Soloviev, "Tight coupling of GPS, laser scanner, and inertial measurements for navigation in urban environments," 2008 IEEE/ION Position, Location and Navigation Symposium, Monterey, CA, 2008, pp. 511-525.
[2] J. Choi, "Hybrid map-based SLAM using a Velodyne laser scanner," 17th International IEEE Conference on Intelligent Transportation Systems (ITSC), Qingdao, 2014, pp. 3082-3087.
[3] Y. Xu, V. John, S. Mita, H. Tehrani, K. Ishimaru and S. Nishino, "3D point cloud map based vehicle localization using stereo camera," 2017 IEEE Intelligent Vehicles Symposium (IV), Los Angeles, CA, 2017, pp. 487-492.
[4] Y. Lu, J. Huang, Y. T. Chen and B. Heisele, "Monocular localization in urban environments using road markings," 2017 IEEE Intelligent Vehicles Symposium (IV), Los Angeles, CA, 2017, pp. 468-474.
[5] F. de Ponte Muller, E. M. Diaz and I. Rashdan, "Cooperative Infrastructure-Based Vehicle Positioning," 2016 IEEE 84th Vehicular Technology Conference (VTC-Fall), Montreal, QC, 2016, pp. 1-6.



Previous Abstract Return to Session D4