Reliability-Aware VPR-Aided Integrated Localization via Evidence Gating and NIS-Based Adaptive Covariance Scaling

Abstract

Reliable localization in Global Navigation Satellite System (GNSS)-challenged environments requires external position updates to constrain inertial drift. Visual place recognition (VPR) can provide such constraints by matching onboard images against a pre-built geo-referenced visual map and generating position pseudo-measurements for integrated localization. However, VPR-derived pseudo-measurements are vulnerable to gross outliers caused by severe illumination changes, seasonal variations, and perceptual aliasing. Directly fusing such unreliable measurements into an inertial navigation system (INS) estimator may produce erroneous state corrections, thereby degrading localization accuracy and filter consistency. To address this challenge, this paper proposes a reliability-aware two-stage measurement management framework for VPR-aided integrated localization. At the retrieval stage, an evidence gate estimates the reliability of each VPR measurement using three complementary cues: descriptor distance, top- $K$ score-distribution uncertainty, and candidate-location dispersion. This gate enables the proactive rejection of high-risk visual pseudo-measurements before fusion. At the filtering stage, an adaptive covariance scaling module uses the normalized innovation squared (NIS) to assess each accepted measurement. When the innovation is statistically inconsistent with the predicted state, the module inflates the measurement covariance to reduce its influence on the filter update. On the Oxford RobotCar dataset, the proposed framework reduces the horizontal RMS and 95th-percentile errors under the challenging day–night condition from 23.71 to 1.86 m and from 55.39 to 3.88 m, respectively. Note to Practitioners—This paper is motivated by practical VPR/INS localization systems operating in GNSS-challenged environments. In such systems, VPR can provide useful map-referenced position updates, but incorrect visual matches may drive the filter estimate toward erroneous locations and cause long-tail localization errors. Therefore, practitioners need a mechanism to determine not only whether a VPR match is available but also whether the resulting pseudo-measurement is sufficiently reliable for fusion. The proposed framework can be integrated into existing VPR/INS systems as a measurement-management module. It first uses retrieval-side evidence to reject unreliable visual pseudo-measurements before fusion. Then, the accepted measurements are further checked by an NIS-based consistency test, with adaptive covariance scaling applied during the filter update. The method operates on standard VPR retrieval outputs without altering the inertial mechanization. The framework is particularly well suited to long-term localization in road vehicles, railway systems, and mobile robots, where changes in illumination, seasonal variations, and perceptual aliasing can lead to visual mismatches.