bevlshaper

Algorithm for bird's-eye-view L-shape fitting in 3D LIDAR point clouds from traffic scenarios

---

🔨 UNDER DEVELOPMENT 🔧

Foreword

This project is inspired by the paper Efficient L-Shape Fitting for Vehicle Detection Using Laser Scanners. As in the paper, a search tree algorithm is used for segmentation after point cloud filtering. This step is inspired by Moving object classification using horizontal laser scan data. Within found clusters, L-shapes are detected. Therefore, rectangles are searched and reduced to L-shapes (oriented towards the sensing vehicle) afterwards. For easy prototyping and modelling, Python and the NumPy library are used instead of a more computationally powerful language.

Prepare environment

conda create --name kitti -y python=3 \
&& conda activate kitti \
&& git clone https://github.com/rwschubert/exploreKITTI.git \
&& pip install numpy pykitti matplotlib opencv-python opencv-contrib-python moviepy \
&& mkdir frames

Run bevlshaper on KITTI dataset scenes

time python bevlshaper.py 2011_09_26 0001

Therefore, the KITTI dataset must have been downloaded in beforehand. In the future, a download script will be supplied

Current issues

• Insufficient distance for recognition
• Insufficient execution speed