AUTONOMOUS VEHICLES LOCALISATION BASED ON SEMANTIC MAP MATCHING METHOD

He Huang, Yu Dong, Jian Yang, Xiaoyi Liu2023 · DOI

TL;DR

This paper introduces a novel semantic map matching method for autonomous vehicle localization, specifically designed for scenarios where GPS signals are unreliable or unavailable. By combining semantic segmentation with visual odometry and a robust map matching strategy, the approach achieves accurate and continuous localization, crucial for safe autonomous driving.

What Problem Does This Paper Address?

Accurate localization is fundamental for autonomous driving systems, enabling vehicles to understand their position and navigate safely. However, in urban environments, GPS signals can be frequently obstructed or interrupted by tall buildings, leading to a loss of positioning accuracy. Traditional methods like GNSS/IMU fusion degrade significantly in such scenarios. This paper addresses the critical need for a robust localization solution that can function reliably even without satellite signals, by leveraging semantic information from camera data to build and match against a semantic map.

Methods

The proposed method involves three main stages:

  • Static Semantic Map Construction: A static semantic map is built using monocular camera data. This process starts with semantic segmentation of road scenes to identify and classify various objects (e.g., buildings, roads, vehicles). Dynamic objects are filtered out to ensure map stability. Feature points are then extracted from static regions, and their geometric constraints are used to refine pose estimation. Finally, these feature points are projected onto a 2D plane to create a semantic map, which is then aligned with the world coordinate system.
  • Semantic Map Matching Localization: During operation, the vehicle builds a local semantic map from its current sensor data. This local map is then matched against the pre-built global semantic map to determine the vehicle's precise position. To overcome the scale ambiguity inherent in monocular vision, the method incorporates a scale recovery technique that uses the camera's mounting height and ground plane fitting to estimate the true translation scale. A semantic constraint ICP (Iterative Closest Point) algorithm is employed for matching, using visual odometry poses as initial estimates.
  • Fusion of Visual Odometry and Map Matching: To enhance accuracy and continuity, the localization results from visual odometry and semantic map matching are fused. This fusion leverages the strengths of both methods: visual odometry provides continuous inter-frame pose estimates, while map matching offers absolute positioning without cumulative error. A sliding window optimization approach is used to combine these information sources for a more stable and accurate final pose.

Datasets

  • CARLA self-driving car simulator: Used for generating training data for the semantic segmentation model.
  • KITTI dataset (00 sequence): Used for evaluating the absolute positioning accuracy of the proposed method.

Key Insights and Ideas

  • Semantic information is key for robust localization: By segmenting the environment and focusing on static semantic features, the method creates a more reliable map and matching process, especially in challenging GPS-denied environments.
  • Handling dynamic objects is crucial: The paper emphasizes the importance of identifying and filtering out dynamic objects from the map to prevent them from negatively impacting localization accuracy and map stability.
  • Monocular scale recovery is essential: The proposed method addresses the inherent scale ambiguity of monocular cameras by integrating ground plane fitting and camera height information, enabling accurate trajectory reconstruction.
  • Fusion of complementary methods yields superior results: Combining the continuous nature of visual odometry with the absolute positioning of semantic map matching provides a more robust and accurate localization solution than either method alone.

Key Results

  • The proposed semantic map matching method demonstrates improved absolute positioning accuracy compared to visual odometry alone on the KITTI dataset.
  • The error distribution of the map matching method is more concentrated, indicating greater stability and reduced cumulative error.
  • The method effectively corrects trajectory offsets and reduces positioning errors, with a median error of 0.47m compared to 0.78m for visual odometry.

Limitations and Future Work

  • Map construction reliance on initial data: The accuracy of the semantic map is dependent on the quality of the initial data and the semantic segmentation performance.
  • Potential for feature point sparsity: In environments with limited distinct static features, the map matching process might still face challenges.
  • Future work could explore:
    • Improving the accuracy of semantic segmentation for more complex scenes.
  • Investigating more advanced fusion techniques for visual odometry and map matching.
  • Extending the method to handle more dynamic environments or different sensor modalities.

Summary updated

Abstract

In autonomous driving systems, a positioning method that can be used in scenarios with no satellite signals or long signal interruptions is a must. In this paper, we address the problems in map construction methods and map matching methods in scenes without satellite signals or long signal interruptions, and construct a semantic map matching-based localization method to meet the localization requirements by means of monocular vision sensors on the basis of weighing the accuracy and cost of map construction and localization. In this paper, firstly, the method of map construction is studied and a static semantic map construction method based on monocular camera is constructed. Then the map matching localization method is studied, and a semantic map matching based localization method is constructed to align the local map built during localization with the pre-built global semantic map to obtain the current location information. Finally, this paper constructs a method to fuse the visual odometry and map matching localization results, so as to obtain more accurate localization results.
BibTeX
@inproceedings{huang2023autonomous,
  author = {He Huang and Yu Dong and Jian Yang and Xiaoyi Liu},
  title = {{AUTONOMOUS VEHICLES LOCALISATION BASED ON SEMANTIC MAP MATCHING METHOD}},
  booktitle = {˜The},
  year = {2023},
  doi = {10.5194/isprs-archives-xlviii-1-w2-2023-901-2023},
}
Figure