Absolute Localization using Visual Data for Autonomous Vehicles

Safa Ouerghi, Rémi Boutteau, Pierre Merriaux, Nicolas Ragot, Xavier Savatier, Pascal VasseurVisigrapp (3: Visapp) 2016 · DOI

TL;DR

This paper presents a novel method for achieving absolute localization of autonomous vehicles using only monocular camera data and a low-cost GPS. By building a compact, geolocalized map of landmarks offline and then using this map for online localization, the system achieves centimeter-level accuracy, a crucial step towards full vehicle automation.

What Problem Does This Paper Address?

Autonomous vehicles require precise localization to navigate safely and effectively, especially as we move towards higher levels of automation (SAE levels 3-5). While advanced sensors like LiDAR offer high accuracy, their cost and complexity are often prohibitive for mass-market adoption. Vision-based localization is therefore highly desirable due to the ubiquity and low cost of cameras. However, traditional vision-based methods often struggle with absolute localization (i.e., knowing the vehicle's position in a global coordinate frame) and can be susceptible to environmental changes. This paper aims to bridge this gap by proposing a robust and cost-effective visual localization system that provides absolute pose information.

Methods

The proposed method consists of two main stages:

  • Map-building (Offline):
    • An instrumented vehicle equipped with an IMU and RTK-GPS drives a route to collect synchronized visual data and precise pose information.
  • Keyframes are selected from the acquired image sequence based on the number of matched keypoints between consecutive frames. This ensures sufficient visual distinctiveness and avoids issues with static scenes.
  • Keypoints (using the FREAK descriptor for its speed and low memory footprint) are detected and matched between consecutive keyframes.
  • These matched keypoints are then triangulated using the precise IMU/GPS poses to generate geolocalized 3D points.
  • The resulting map is a compact database containing keyframes, their associated keypoints, descriptors, and the 3D coordinates of these points. This map is significantly smaller than the raw image data.
  • Localization (Online):
    • This stage does not require the expensive IMU/RTK-GPS system, relying only on a monocular camera and a low-cost GPS.
  • The low-cost GPS is used to find the nearest keyframe in the pre-built map, providing an initial guess for the vehicle's position.
  • Keypoints are detected in the current camera frame and matched against the keypoints of the nearest keyframe.
  • Outliers are rejected using a RANSAC scheme with the Eight-Point algorithm (chosen for its balance of speed and accuracy in this context).
  • The 2D-3D correspondences between the current image and the map are then used to solve the Perspective-n-Point (PnP) problem, yielding the absolute pose of the vehicle.

Datasets

  • KITTI Vision Benchmark Suite: This dataset was used for both map-building and evaluation. The authors specifically mention using images from the left grayscale camera, along with IMU/RTK GPS ground truth poses and camera calibration matrices.

Key Insights and Ideas

  • Offline Map-building for Online Efficiency: The separation of map-building (offline, with high-precision sensors) and localization (online, with low-cost sensors) is a key design choice that enables practical deployment.
  • Keyframe Selection for Compactness and Robustness: The intelligent selection of keyframes, rather than using every frame, drastically reduces map size while ensuring sufficient visual parallax for accurate 3D reconstruction.
  • FREAK Descriptors for Speed and Memory: The choice of FREAK descriptors is highlighted for its efficiency, making the online localization process faster and less memory-intensive.
  • Absolute Localization with Low-Cost Sensors: The core contribution is demonstrating that centimeter-level absolute localization is achievable using only a monocular camera and a low-cost GPS, which is a significant step towards making advanced autonomous driving features more accessible.

Key Results

  • Map Size Reduction: The built map was approximately 40 times smaller than the original KITTI database, with keyframes spaced about 3 meters apart on average, compared to 80 cm in the original dataset.
  • Localization Accuracy: The system achieved a mean localization error of around 17 centimeters, which the authors deem sufficient for autonomous vehicle control.
  • Computational Efficiency: The Eight-Point algorithm for outlier rejection was significantly faster (11 ms) than the Five-Point algorithm (52 ms) while maintaining a comparable proportion of inliers.

Limitations and Future Work

  • Robustness to Environmental Changes: The authors acknowledge that changes in lighting, seasons, or the appearance/disappearance of objects between map acquisition and localization could affect performance.
  • Future Work: They suggest improving robustness by incorporating multiple feature fusion techniques to handle these environmental variations. They also mention that discontinuities in estimated poses could be smoothed using a Kalman Filter, and the reprojection error of 3D points could be used as a measure of estimation confidence.

Summary updated

BibTeX
@inproceedings{ouerghi2016absolute,
  author = {Safa Ouerghi and Rémi Boutteau and Pierre Merriaux and Nicolas Ragot and Xavier Savatier and Pascal Vasseur},
  title = {{Absolute Localization using Visual Data for Autonomous Vehicles}},
  booktitle = {Visigrapp (3: Visapp)},
  year = {2016},
  doi = {10.5220/0005682105950601},
}
Figure