EROAM: Event-based Camera Rotational Odometry and Mapping in Real-time

IMPORTANT NOTE: This code is released primarily for the purpose of academic review, as the accompanying paper is currently under peer review. We welcome feedback and suggestions.

EROAM System Overview

EROAM is a novel, real-time event-based system for accurate camera rotational odometry and mapping. It addresses the limitations of traditional methods in high-speed scenarios by uniquely combining several key innovations:

• Core Principle: Instead of relying on event generation models or contrast maximization, EROAM introduces a Spherical Event Representation. Events are projected onto a unit sphere, simplifying rotation math and enabling continuous, high-resolution mapping.
• Estimation Algorithm: The system features the Event Spherical Iterative Closest Point (ES-ICP) algorithm, a geometric optimization framework specifically designed to efficiently align sparse event point clouds on the sphere using robust point-to-line metrics.
• High-Frequency & Real-time: EROAM operates at high frequencies (e.g., 1000 Hz) entirely on CPU. This ensures robust tracking of dynamic motions and provides excellent initial guesses for ES-ICP, contributing to its accuracy and real-time capability.
• Efficient Long-Term Mapping: For map management, EROAM employs an incremental k-d tree (ikd-Tree) and Regional Density Management (RDM). This allows for stable computational costs during extended operation and prevents map degradation. The system architecture comprises a Tracking Module for high-frequency pose estimation and a Mapping Module for maintaining an efficient global spherical map.
• Output: The system delivers accurate $SO(3)$ pose estimates and generates high-quality panoramic images with fine structural details at arbitrary resolutions as a byproduct.

Publication

Our paper is currently available on arXiv and is under peer review. If you find EROAM useful in your work, we would appreciate a citation to our arXiv preprint:

@article{xing2024eroam,
  title={EROAM: Event-based Camera Rotational Odometry and Mapping in Real-time},
  author={Xing, Wanli and Lin, Shijie and Yang, Linhan and Zhang, Zeqing and Du, Yanjun and Lei, Maolin and Pan, Yipeng and Pan, Jia},
  journal={arXiv preprint arXiv:2411.11004},
  year={2024}
}

You can find more details on our Project Page and read the preprint on arXiv.

Video

We have created comprehensive single-take video demonstrations available at this YouTube link. These feature 12 distinct segments that showcase EROAM running in real-time across a wide spectrum of challenging scenarios, including static scenes, wide-range rotations, moving pedestrians, vehicles, motorcycles, pause-and-resume operations, night scenes, and the LimX TRON 1 robot platform. These demonstrations simultaneously capture the event camera in motion, the real-world scene, and the algorithm’s real-time output on a laptop screen.

Datasets

We provide the EROAM-campus dataset used in our paper, which includes real-world sequences with event camera data and LiDAR-based ground truth for rotational motion. You can download it from this SharePoint link.

Prerequisites

Ensure the following dependencies are installed:

• ROS1 Noetic (Installation Guide)
• iniVation DV ROS driver (dv-ros)
• Eigen3
• Sophus
• Glog
• OpenCV
• PCL
• libtbb-dev
• (And other standard build tools like CMake, g++)

Installation

1. Set up your Catkin Workspace:

   mkdir -p ~/eroam_ws/src
   cd ~/eroam_ws/src

2. Clone Repositories: Clone the iniVation DV ROS driver and the EROAM source code into your workspace's src directory.

   git clone https://gitlab.com/inivation/dv/dv-ros.git
   git clone git@github.com:wlxing1901/eroam.git

3. Build the Project: Navigate to your workspace directory and use catkin_make.

   cd ~/eroam_ws
   catkin_make

4. Source the Workspace: After a successful build, source the setup.bash file.

   source ~/eroam_ws/devel/setup.bash

   It's recommended to add this line to your .bashrc for convenience.

Usage

1. Download Data: Download the eroam_data from this SharePoint link. This link contains the datasets used for testing.

2. Configure Launch File: Open the eroam_run_campus.launch file located in the eroam package (e.g., ~/eroam_ws/src/eroam/launch/eroam_run_campus.launch). Modify the following line to point to the correct path of your downloaded rosbag data:

   <node name="rosbag_play" pkg="rosbag" type="play" args="-d 3.0 path_to_your_data"/>

   Replace path_to_your_data with the actual path to your .bag file. For example:

   <node name="rosbag_play" pkg="rosbag" type="play" args="-d 3.0 /path/to/your/downloaded/dataset.bag"/>

3. Run EROAM: Execute the launch file:

   roslaunch eroam eroam_run_campus.launch

4. Visualization and Termination: The program will launch RViz to visualize the EROAM execution. Wait for the rosbag to finish playing and for EROAM to complete processing. Once processing is done, press Ctrl+C to terminate the program.

   Note: RViz visualization might experience lag if the accumulated point cloud in the map becomes very dense. This typically does not affect the core EROAM algorithm's performance.

5. Output and Performance Metrics: After termination, the terminal will display performance metrics, including the number of calls and average execution time for key steps:

   I20250609 17:31:00.751452  8102 eroam.hpp:371] EROAM: done.
   I20250609 17:31:00.751473  8102 timer.cc:12] >>> ===== Printing run time =====
   I20250609 17:31:00.751477  8102 timer.cc:14] > [ EROAM: ev_sphere_frame warp ] average time usage: 0.060079 ms , called times: 36274
   I20250609 17:31:00.751510  8102 timer.cc:14] > [ EROAM: event spherical icp ] average time usage: 0.585088 ms , called times: 36267
   I20250609 17:31:00.751541  8102 timer.cc:14] > [ EROAM: full time ] average time usage: 0.713804 ms , called times: 41081
   I20250609 17:31:00.751578  8102 timer.cc:14] > [ EROAM: ikd-Tree initial build ] average time usage: 1.00224 ms , called times: 1
   I20250609 17:31:00.751593  8102 timer.cc:14] > [ EROAM: ikd-tree update ] average time usage: 0.38847 ms , called times: 1683
   I20250609 17:31:00.751598  8102 timer.cc:14] > [ EROAM: map_frame_world copy ] average time usage: 0.00113396 ms , called times: 1683
   I20250609 17:31:00.751605  8102 timer.cc:14] > [ EROAM: pubish everything ] average time usage: 0.0636827 ms , called times: 36267
   I20250609 17:31:00.751636  8102 timer.cc:14] > [ EROAM: saving panoramic image ] average time usage: 5469.19 ms , called times: 1
   I20250609 17:31:00.751639  8102 timer.cc:14] > [ EROAM: saving pose result ] average time usage: 80.9835 ms , called times: 1
   I20250609 17:31:00.751642  8102 timer.cc:14] > [ EROAM: undistortion and projection ] average time usage: 0.0251254 ms , called times: 36274
   I20250609 17:31:00.751672  8102 timer.cc:19] >>> ===== Printing run time end =====
   

   The panoramic image and estimated pose trajectory will be saved to an output directory (e.g., ~/eroam_ws/src/eroam_opensource_test/output/ or as specified in your launch/config files).

Contact

For questions or support, please open an issue on GitHub or contact wlxing@connect.hku.hk