Introduction

Ability to detect cars in any KITTI rosbag using PointRCNN detections. Able to visualize in image and LiDAR space.

Requirements

• Ubuntu 18.04
• Pytorch 1.0
• ROS Melodic
• Python 3+
• Rosbag
• Pre - trained model

Prepare environment

It is recommended to create a virtual environment.

python3 -m venv path/to/virtual/env/name

source path/to/virtual/env/name/bin/activate

Create catkin workspace.

mkdir -p ~/catkin_ws/src

cd ~/catkin_ws/src

Clone repo, and include the submodules

git clone --recursive https://github.com/dmatos2012/pointrcnn_detector_ros.git

Once you have the venv setup, install python 3 requirements, in the root folder(/pointrcnn_detector_ros)

pip install -r requirements.txt

Additionally, you need to build extra PointRCNN libraries (iou3d, pointnet2, roipool3d)

cd src/PointRCNN/

sh build_and_install.sh

Then run the script to build the environment with cv bridge and python3.

cd ~/catkin_ws

sh src/pointrcnn_detector_ros/configure_environment.sh path/to/virtual/env/name/bin/python

source devel/setup.bash

Data Preparation

• ROSBAG: In order to convert a KITTI raw data to a rosbag, check out kitti2bag. Although you can download the one I used here.

  Place the rosbag in pointrcnn_detector_ros/data/kitti_2011_09_26_drive_0011_synced.bag

  For now, I have hardcoded the launch of this specific rosbag, but you can just modify the roslaunch file and change it to your own rosbag. If you use a different rosbag, make sure you also change the calibration file to match that specific rosbag.

• Pre-trained model: You can obtained pre - trained model from the PointRCNN repo here.

  Place the pre-trained model in pointrcnn_detector_ros/src/PointRCNN/tools/PointRCNN.pth

Visualize

Launch visualization with

roslaunch prcnn_ros kitti_car_det.launch

Acknowledgements

PointRCNN for the amazing object detector and their publicly available pre-trained model and code.

Kitti2bag for tool to convert KITTI raw data to rosbag