Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Evaluating Visual Odometry Methods for Autonomous Driving in Rain

Yu Xiang Tan, Marcel Bartholomeus Prasetyo, Mohammad Alif Daffa, Deshpande Sunny Nitin, Malika Meghjani

TL;DR

The paper tackles the challenge of visual odometry reliability for autonomous urban driving in rainy conditions. It conducts a comprehensive, rain-focused evaluation of seven open-source VO algorithms plus a DROID-SLAM based heuristic across diverse rain datasets (Oxford Robotcar, 4Seasons, and an internal Singapore collection) for both monocular and stereo setups, using Absolute Trajectory Error as the primary metric. Key contributions include a systematic comparison of classical, learning-based, and mixed VO approaches in rain, analysis of their strengths and limitations, and the introduction of a DROID-SLAM variant with Conservative Global Reference Path and confidence-driven keyframe strategy to enhance rain robustness. The findings indicate that while depth-prediction-based methods can preserve scale under rain and stereo approaches offer some long-term localization benefits, none of the VO methods alone are sufficient in rain, underscoring the need for sensor fusion to enable robust autonomous operation in adverse weather.

Abstract

The increasing demand for autonomous vehicles has created a need for robust navigation systems that can also operate effectively in adverse weather conditions. Visual odometry is a technique used in these navigation systems, enabling the estimation of vehicle position and motion using input from onboard cameras. However, visual odometry accuracy can be significantly impacted in challenging weather conditions, such as heavy rain, snow, or fog. In this paper, we evaluate a range of visual odometry methods, including our DROID-SLAM based heuristic approach. Specifically, these algorithms are tested on both clear and rainy weather urban driving data to evaluate their robustness. We compiled a dataset comprising of a range of rainy weather conditions from different cities. This includes, the Oxford Robotcar dataset from Oxford, the 4Seasons dataset from Munich and an internal dataset collected in Singapore. We evaluated different visual odometry algorithms for both monocular and stereo camera setups using the Absolute Trajectory Error (ATE). From the range of approaches evaluated, our findings suggest that the Depth and Flow for Visual Odometry (DF-VO) algorithm with monocular setup performed the best for short range distances (< 500m) and our proposed DROID-SLAM based heuristic approach for the stereo setup performed relatively well for long-term localization. Both VO algorithms suggested a need for a more robust sensor fusion based approach for localization in rain.

Evaluating Visual Odometry Methods for Autonomous Driving in Rain

TL;DR

The paper tackles the challenge of visual odometry reliability for autonomous urban driving in rainy conditions. It conducts a comprehensive, rain-focused evaluation of seven open-source VO algorithms plus a DROID-SLAM based heuristic across diverse rain datasets (Oxford Robotcar, 4Seasons, and an internal Singapore collection) for both monocular and stereo setups, using Absolute Trajectory Error as the primary metric. Key contributions include a systematic comparison of classical, learning-based, and mixed VO approaches in rain, analysis of their strengths and limitations, and the introduction of a DROID-SLAM variant with Conservative Global Reference Path and confidence-driven keyframe strategy to enhance rain robustness. The findings indicate that while depth-prediction-based methods can preserve scale under rain and stereo approaches offer some long-term localization benefits, none of the VO methods alone are sufficient in rain, underscoring the need for sensor fusion to enable robust autonomous operation in adverse weather.

Abstract

The increasing demand for autonomous vehicles has created a need for robust navigation systems that can also operate effectively in adverse weather conditions. Visual odometry is a technique used in these navigation systems, enabling the estimation of vehicle position and motion using input from onboard cameras. However, visual odometry accuracy can be significantly impacted in challenging weather conditions, such as heavy rain, snow, or fog. In this paper, we evaluate a range of visual odometry methods, including our DROID-SLAM based heuristic approach. Specifically, these algorithms are tested on both clear and rainy weather urban driving data to evaluate their robustness. We compiled a dataset comprising of a range of rainy weather conditions from different cities. This includes, the Oxford Robotcar dataset from Oxford, the 4Seasons dataset from Munich and an internal dataset collected in Singapore. We evaluated different visual odometry algorithms for both monocular and stereo camera setups using the Absolute Trajectory Error (ATE). From the range of approaches evaluated, our findings suggest that the Depth and Flow for Visual Odometry (DF-VO) algorithm with monocular setup performed the best for short range distances (< 500m) and our proposed DROID-SLAM based heuristic approach for the stereo setup performed relatively well for long-term localization. Both VO algorithms suggested a need for a more robust sensor fusion based approach for localization in rain.
Paper Structure (30 sections, 4 figures, 2 tables)

This paper contains 30 sections, 4 figures, 2 tables.

Table of Contents

  1. INTRODUCTION
  2. RELATED WORK
  3. Localization in Adverse Weather
  4. Utilizing Robust Sensors
  5. Sensor Fusion Approaches
  6. Vision in Adverse Weather
  7. Robust Feature Detectors
  8. Removing Noise Artefacts
  9. Visual Odometry
  10. Direct vs Indirect Approaches
  11. Dense vs Sparse Approaches
  12. Learning-based vs Classical Approaches
  13. Mixed Approaches
  14. EVALUATED VISUAL ODOMETRY ALGORITHMS
  15. Direct Sparse Odometry (DSO)
  16. ...and 15 more sections

Figures (4)

  • Figure 1: Sample images from the three datasets.
  • Figure 2: Comparison of blur value for the same scene taken from the $11$-$21$ and $12$-$09$ sequences respectively, from the Oxford Robotcar Dataset
  • Figure 3: Output trajectories of DF-VO compared with SVO evaluated on the entire route of the 05-29 sequence from the Oxford Robotcar Dataset. DF-VO using a depth prediction model has a more consistent scale across the route while SVO has inconsistent scale as shown by the varying estimated path lengths compared to the ground truth.
  • Figure 4: TartanVO evaluation output on the first 500m of the 10-29 sequence from the Oxford Robotcar dataset. The red line shows the segment of the route where the vehicle is not moving in the video.