FLOW-VINS

Introduction

This project is developed based on VINS-Fusion to solve the visual SLAM problem in dynamic scenes. This is also my graduation project

The work that has been completed so far is：

add a instance segmentation model ( Yolov8n-seg ), to detect dynamic objects in scene, after acceleration by TensorRT, segment speed is reach 10ms each image(GTX 1070).
make a hierarchical management of feature points: Limit the scale of back-end optimization by classifying feature points and improve the speed of back-end optimization， by test, the optimization speed is up to 100Hz each iterations(11th i5)
Break away from the ROS framework as much as possible (but not completely), and merge multiple nodes into a single node.

Work to be done is：

Adding a magnetometer to overcome the heading drift of the VIO system.
Increase the health detection mechanism and adjust the weight of sensor data during backend optimization online.

Environment

ROS Noetic
CUDA 11.2 (for Segment)
TensorRT 8.2.5.1 (for Segment)
OpenCV 4.7.0 (for Segment)
Ceres Solver
Eigen

Example

Clone the repository and catkin_make:

    cd ~/catkin_ws/src
    git clone https://github.com/ErenJaeger-01/Flow-VINS.git
    cd ../
    catkin_make
    source ~/catkin_ws/devel/setup.bash

EuRoC Example

（Loop detection is currently enabled by default）

Monocular camera + IMU

    roslaunch vio_system vins_euroc_mono_imu.launch
    
    rosbag play YOUR_DATASET_FOLDER/MH_01_easy.bag

Stereo camera + IMU

    roslaunch vio_system vins_euroc_stereo_imu.launch
    
    rosbag play YOUR_DATASET_FOLDER/MH_01_easy.bag

Stereo camera

    roslaunch vio_system vins_euroc_stereo.launch
    
    rosbag play YOUR_DATASET_FOLDER/MH_01_easy.bag

TUM Dynamic Example

    roslaunch vio_system vins_tum_rgbd.launch
    
    rosbag play YOUR_DATASET_FOLDER/rgbd_dataset_freiburg3_walking_xyz.bag

Kaixiang-Z/Flow-VINS