/VoxelNetRos

implement the VoxelNet with ROS, using Kitti data to test

Primary LanguagePython

Voxelnet ROS Implementation

VoxelNet Implementation codes from "https://github.com/qianguih/voxelnet"

Dependencies

  • python3.5+
  • TensorFlow (tested on 1.4)
  • opencv
  • shapely
  • numba
  • easydict
  • ROS(melodic)
  • jsk package(ros-jsk-recognition)
  • cv_brdge(ros)
  • nodelet(ros)
  • pcl_ros(ros)
  • rviz(ros)

Data Preparation

Download the test data from the raw KITII dataset

unzip to data folder,Directory structure should be:

data
----npy_lidar_0048(lidar point cloud)
----result_road_0048(road detection result)
----rgb_image_0048(correponding rgb left camera image)

Instructions

before run the code, you may need to install:

  • clone this repository
  • move voxelnet_ros folder to your catkin_ws
  • catkin_make
  • roscd voxelnet/script/
  • you probably also need to complie the box_overlap.so using numba
  • python3 voxelnet_ros.py & python3 pub_kitti_point_cloud.py
    • unfortunately, rosrun won't work. because it's using Python 3.x instead of 2.x
  • roslaunch voxelnet_ros convex_hull.launch
  • rosrun rviz rviz
    • use the voxelnet.rviz to config rviz

ROS Node

Rviz Animation

Future Work

  • Retrain the model
  • Still too slow, need more efficient implementation for the VoxelNet
  • code clean