/multisensor-fusion-for-localization-and-mapping

A framework for multisensor fusion localization and mapping

Primary LanguageC++

Multisensor Fusion for Localization and Mapping

This project aims to provide a open-sourced framework for multisensor fusion in localization and mapping.
Install the following dependencies to run this project:

  1. ROS
  2. g2o
  3. GeographlicLib
  4. glog
  5. PCL
  6. protobuf
  7. sophus
  8. YAML

Dataset

This is a modified kitti rosbag: https://drive.google.com/file/d/1zoc7hm5nZvv0koZVPxt19J5d-TOxVfK1/view?usp=sharing
The original kitti dataset has many missing parts in the raw data. The missing part has been linearly interpolated in this rosbag for the purpose of multisensor fusion. However, linear interpolation does not reflect how the actual missing data are like.

Multisensor Mapping

This package implements multisensor mapping using LiDAR, IMU, GNSS, and odometer measurements to build a map and evaluates performance on the KITTI dataset.
The frontend module uses NDT CPU from Autoware to provide LiDAR poses. The loop closure module uses Scan Context to detect loop closures. The backend fuses LiDAR relative poses, IMU and odometer preintegrations, GNSS prior positions, and loop closure relative poses as optimization constraints. Since mapping does not need to run in real-time, a sliding window is not implemented here.

Map

map

LiDAR Only Mapping

Multisensor Fusion Mapping

The experiments show obvious improvement of multisensor fusion on mapping accuracy.

TODO

  1. Integrate gtsam-based factor graph optimization as mapping backend.

Multisensor Localization

This package implements multisensor localization using LiDAR, IMU, GNSS, and odometry measurements to localize within a prebuilt map and evaluates performance on the KITTI dataset.
The frontend uses NDT CPU from Autoware to provide LiDAR poses and adds Scan Context to detect loop closures. The backend fuses LiDAR relative poses after loop closures, IMU preintegrations, and GNSS prior poses as optimization constraints. A sliding window is implemented to margalize old measurements as prior and speed up optimization.

Trajectory

trajectory

LiDAR Only Localization

Multisensor Fusion Localization

The experiments show obvious improvement of multisensor fusion on localization accuracy.

TODO

  1. Integrate Kalman Filter, Extended Kalman Filter, or Error-State Kalman Filte as localization backend.
  2. Integrate gtsam-based factor graph optimization as localization backend.

References

[1] Autoware Foundation. (2020). NDT_CPU. https://github.com/autowarefoundation/autoware_ai_perception/tree/master/ndt_cpu.
[2] Qian Ren. (2020). Localization in Auto Driving. https://github.com/Little-Potato-1990/localization_in_auto_driving.
[3] Tong Qin. (2019). VINS-Mono. https://github.com/HKUST-Aerial-Robotics/VINS-Mono.
[4] Yao Ge. (2021). Sensor Fusion for Localization CourseWare. https://github.com/AlexGeControl/Sensor-Fusion-for-Localization-Courseware.