Integration of Visual SLAM into Consumer-Grade Automotive Localization

TL;DR

The paper addresses the challenge of accurate ego-motion estimation in consumer-grade vehicles by integrating visual SLAM into a proprioceptive localization pipeline to enable online gyroscope calibration. It introduces a lateral-velocity vehicle model and an adaptive Kalman filter to jointly estimate state and gyroscope-related parameters, leveraging camera measurements and wheel encoders. The approach yields precise online calibration (offsets within ≈0.05 deg/s and yaw-scale within ≈0.1%) and improves localization accuracy, outperforming state-of-the-art SLAM methods on public benchmarks and maintaining benefits even when vision is temporarily unavailable. This work offers a practical pathway to higher automotive localization accuracy by fusing visual SLAM with proprioceptive sensing in a robust, calibration-friendly framework.

Abstract

Accurate ego-motion estimation in consumer-grade vehicles currently relies on proprioceptive sensors, i.e. wheel odometry and IMUs, whose performance is limited by systematic errors and calibration. While visual-inertial SLAM has become a standard in robotics, its integration into automotive ego-motion estimation remains largely unexplored. This paper investigates how visual SLAM can be integrated into consumer-grade vehicle localization systems to improve performance. We propose a framework that fuses visual SLAM with a lateral vehicle dynamics model to achieve online gyroscope calibration under realistic driving conditions. Experimental results demonstrate that vision-based integration significantly improves gyroscope calibration accuracy and thus enhances overall localization performance, highlighting a promising path toward higher automotive localization accuracy. We provide results on both proprietary and public datasets, showing improved performance and superior localization accuracy on a public benchmark compared to state-of-the-art methods.

Figures (6)

• Figure 1: On-board camera tracking features in an urban environment. Green indicates successful tracking, with the yellow ellipses showing the uncertainty of the track.
• Figure 2: Overview of the proposed method. The IMU measurements are used to propagate both the vehicle motion and the relative feature motion. The vehicle's velocity measurements and the camera images are fed through the measurement model. The vehicle-velocity model also includes our lateral-velocity model. Within the adaptive Kalman filter framework the state and uncertainty are predicted and updated accordingly. The feedback loops are omitted for visual clarity.
• Figure 3: Estimated offsets $\mathbf{b}_{\omega}$ over time with their ground-truth values.
• Figure 4: Estimated yaw-rate scale error $s_{\omega_z}$ over time (above). Lower plot depicts the deviation from the ground-truth value over the GNSS position.
• Figure 5: Estimated misalignment parameters $s_{\omega_{yx}}$ and $s_{\omega_{xy}}$ and their reference value.