By Urs Niesen, Jubin Jose, Xinzhou Wu, Qualcomm Technologies Inc.
Emerging automotive applications require reliable but at the same time low-cost positioning solutions. In this paper, we present such a solution by fusing the measurements from several consumer-grade sensors using a tightly coupled centralized filter.
The sensors used are a single-frequency GNSS receiver providing GPS and GLONASS pseudoranges and GPS carrier-phase measurements, a micro-electro-mechanical (MEMS) inertial measurement unit (IMU), a monocular camera, wheel-speed and steering-angle sensors.
We also employ vehicular constraints, integrated as pseudo-measurements. The centralized fusion architecture allows sensor cross-calibration and improves outlier detection. The filter runs in real time on the target platform, producing pose estimates at 30 Hz. Through extensive experimental evaluations, we demonstrate positioning accuracies of sub-meter 95-percentile horizontal errors even in GNSS-challenged deep-urban scenarios.
Conflicting Requirements. Accurate positioning is a requirement for several emerging vehicular applications such as advanced driver-assistance systems (ADAS) and autonomous driving. Positioning solutions for these applications face two competing constraints. To be technically viable, the computed position estimate needs to be reliable in scenarios ranging from open sky to deep urban, with less than 1-meter 95-percentile horizontal error as an often-mentioned target. To be economically viable, the system needs to be built from consumer-grade components.
We reconcile these conflicting requirements by fusing measurements from several low-cost sensors into a single pose estimate using one centralized extended Kalman filter (EKF). A multi-constellation single-frequency GNSS receiver provides GPS pseudorange and carrier-phase measurements and GLONASS pseudorange measurements. These are combined in a tightly coupled integration architecture with a consumer-grade MEMS IMU used to produce the reference navigation solution.
Tight integration enables outlier rejection directly for the raw GNSS measurements. This is crucial in deep-urban scenarios, since many or most raw GNSS measurements could be outliers in these conditions. We use a monocular camera and vehicular sensors, providing four wheel-speed measurements and a steering-angle measurement, as additional aiding sensors.
Constraints. Finally, vehicular constraints are integrated as pseudo-measurements. These sensors have very different noise sources and failure modes, which allows cross-calibration and improves failure and outlier detection. Given the tightly coupled integration in a single EKF, the filter state is quite large and can reach more than 100 dimensions. Despite its size, we are able to run the filter in real time and on target, producing pose outputs at a rate of 30 Hz.
We report the result of extensive experimental evaluations in different scenarios ranging from open sky with good satellite visibility to deep urban with long stretches of no or only limited satellite visibility. In each of these scenarios, we obtain the target accuracy of sub-meter 95% horizontal positioning error.
We show that, in the benign open-sky scenarios, GPS and IMU sensors are sufficient to achieve the target accuracy. However, in challenging deep-urban scenarios, all the integrated sensors are required to attain reliable sub-meter positioning performance.
Sensors and Components. We use Qualcomm SiRFstarV 5e B02 GNSS chipset, a low-cost commercial GNSS product, connected to a NovAtel GPS-702-GG dual-frequency GPS+GLONASS Pinwheel antenna, the only component not consumer-grade, to separate impact of a specific antenna on performance. We plan to evaluate low-cost antennas in the future. We use a TDK InvenSense low-cost MEMS 6-axis IMU (MPU-6150) and a vehicle interface with vehicle sensors through the controller area network bus. Accurate timestamping for tightly coupling sensor measurements is provided by a custom sensor sync board. The processor is a Qualcomm Snapdragon 820 automotive platform for real-time computation. (Qualcomm SiRFstar and Qualcomm Snapdragon are products of Qualcomm Technologies, Inc. and/or its subsidiaries.)
This paper was presented at ION-GNSS+ 2018.
.
