vru-intersection-safety

This is the official repository for the SAE 24 paper Evaluating Safety Metrics for Vulnerable Road Users at Urban Traffic Intersections Using High-Density Infrastructure LiDAR System.

The scripts were successfully tested on an Ubuntu 22.04 system with Intel i9 13th Gen processor, NVIDIA RTX 4090 GPU and 500 GB of disk space.

Project resources can be found at this link.

setup conda env

Create conda environemt with Python 3.10.6 and use pip_list.txt and conda_list.txt as reference for required packages and their versions.

install MMDetection3D

Install mmdet3d following this link.

install ROS2

Install ROS2 following this link. Install conda install -c conda-forge libstdcxx-ng in your conda env to make it compatible with rclpy.

download the PV-RCNN checkpoint

Download the checkpoint from this and move it inside scripts/mmdet3d/

setup local environment

Create Datasets folder inside home.

mkdir -p ~/Datasets/Hesai/

Move hesai_mill_ave_1 and hesai_mill_ave_4 folders from resources inside ~/Datasets/Hesai/ folder. Only these 2 out of 4 datasets have been released.

run detection and tracking

Tracking results can be genertated using the following script. This will generate a file tracking_results.json. Precomputed results are available in resources folder for all the 4 datasets.

# run from project root folder. specify your user in the script before running.
python mmdet3d_pipeline.py

# visualize tracklets in 3D pointcloud.
rviz2 -d scripts/rviz/hesai_viz.rviz

Specific entity interactions and metrics evaluation can be visualized using the following script. In the previous step, look at the rviz screen to point out tracking IDs of two entities of interest and specify them in the script to visualize their interaction.

# run from project root folder. specify your user in the script before running. specify data_id (1 for millave_1 and 4 for millave_4) in the script before running. keep the rviz screen open to see the interaction in 3D pointcloud.
python safety_metrics_viz.py

Quantitative results for each dataset can be generated from the tracking results using the following script.

# run from project root folder. specify your user in the script before running.
python safety_metrics_stats.py

note

This repository only contains the offline scripts supporting the results showcased in the paper. Online metrics evaluation scripts will be released in future.