Skip to content

Experimental platform to achieve fused environment perception using different modalities of sensors

Notifications You must be signed in to change notification settings

Eagleflag88/MultiModalEvnPercption

Repository files navigation

MultiModalEvnPercption

A ROS-based experimental software platform for Multi-Modal Simultaneous Localization And Mapping.

For the purpose of studying and research in the field of SLAM, based on the infrastructure of ROS, this repo provide a software platform which is highly modularized and is able to fuse the information from a number of different modalities to achieve to an accurate and consistent environmental understanding.

Software Architecture

The whole package of software consists of three layers, as depicted in following picture:

Software Architecture

Sensors

Modeling of the sensor driver, which acts as the source of sensor measurements. In the case of simulation, these nodes take the playback of rosbag or other compliant data format.

Frontend

Responsible for preprocessing of the measurements, such as filtering and undistortion. As next steps, feature extraction and motion estimation are implemented. Meanwhile, objection detection based on deep learning is implemented for object-level environmental understanding. In addition, a tracker node is exploited to take care of data association, in feature and object-level.

Backend

While the frontend focuses on the real-time performance, the nodes of the backend try to fuse the initial guesses from frontend to obtain more accurate estimate of the ego poses and landmarks. To this end, a sliding window is maintained to constrain the computational burden and complexity. Within the sliding window, maximum-a-posteriori estimates are calculated based on the inference of a factor graph. Alternatively, an EKF can be applied.

Dependencies

Basic Software Environment

OS: Ubuntu 20.04
ROS: Noetic
OpenCV & Opencv_Contrib: 4.4.0 (For Compatibility with Cuda 11), Consult the Installation Instruction
OpenMP:

sudo apt install libomp-dev    

Software for DL Based Environment Perception

Cuda: 11: Consult the Installation Instruction
Cudnn: 8: Consult the Installation Instruction
TensorRT: Consult the Installation Instruction
Darknet For Yolo: Consult the Installation Instruction

SLAM Related

Ceres: Consult the Installation Instruction; SuiteSparse and CXSparse are included
Eigen 3:

sudo apt install libeigen3-dev

g2o: Comes with the repo, github commit a48ff8c
Geographic_Lib: A library dealing with geographic coordinate conversion, comes with the repo

Building

Clone the Repo

git clone https://github.com/Eagleflag88/MultiModalEvnPercption.git

Build the Ros Package

./build.sh

Todo

  1. Refactor the codebase.
  2. Implement the keyframe_node, factor_graph_node and ekf_node.
  3. Create launch files for different functional modes, for example, camera_only, lidar_only, visual inertial, etc.
  4. Add the node for place recognition into the software architecture.
  5. Add the HD-Map as a further source of information.
  6. Add semantic pipeline

About

Experimental platform to achieve fused environment perception using different modalities of sensors

Topics

Resources

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published

Languages