This repository provides the code for the paper: Drivence: Realistic Driving Sequence Synthesis for Testing Multi-Sensor Fusion Perception Systems [Website]
2026-07-08: 🚀 We have published OpenFusion, a dedicated, operatorized multimodal fusion library that provides comprehensive support and pre-trained weights for multiple mainstream fusion perception models (e.g., BEVFusion, CLOCs, HMFI, and LoGoNet). Check it out! ⭐
Drivence is a realistic driving sequence synthesis tool designed for testing multi-sensor fusion (MSF) perception systems. It can generate realistic driving sequences under various scenarios and evaluate the robustness of MSF perception systems in different traffic conditions. The framework consists of the following components:
Drivence introduces an occupancy grid mapping module that generates an occupancy grid map from input point cloud data of the driving sequence.
The generated map can download here.
Link: https://pan.baidu.com/s/10M6Byp4hFz1YCBRm2ytttg?pwd=mapp
Code: mappDrivence features a three-step trajectory generation process:
- A clustering method identifies waypoints representing all possible driving lanes within the drivable area.
- A global path planner generates a global reference trajectory that aligns with the identified lanes and follows a specified driving pattern.
- A local path planner dynamically adjusts this reference trajectory in real-time to avoid collisions with existing NPC vehicles.
Drivence leverages a physics-aware multi-sensor simulation module to insert multiple types of objects into the driving sequence. This module can simulate data from various sensors, including LiDAR and cameras, to generate realistic sensor data for testing MSF perception systems.
Drivence utilizes metamorphic relations (MRs) to develop test oracles that automatically assess the system’s performance.
Drivence is implemented using PyTorch 1.8.0 and Python 3.7.11. All experiments are conducted on a server equipped with an Intel i7-10700K CPU (3.80 GHz), 48 GB RAM, and an NVIDIA GeForce RTX 3070 GPU (8 GB VRAM).
To install the required dependencies, run:
pip install -r requirements.txt
python build_script.pyNote that the following version numbers must be met, otherwise errors may occur
cuda==11.8
open3d==0.17.0(Error when open3d==0.19.0 is known)
Install Blender DRIVENCE leverage blender, an open-source 3D computer graphics software, to build virtual camera sensor.
- install blender>=3.4.1 from this link
To reproduce our experiments, the environment must be carefully configured for each system:
- For DFmot and JMODT, refer to the benchmarking code in the
AI-MSF-Benchmarkrepository. - For YONTD, follow the instructions in the YONTD repository.
- For VirTrack, follow the instructions in the VirTrack repository.
We provide extensive visualization results. Please refer to our website for more visualization results.
-
Run map generation
cd map_gen python main.py -
Run trajectory generation
cd TrackGen python main.py -
Run sensor simulation
cd data_gen python main_function_new.pyRun
main.pyto generate the driving sequence and sensor data. -
Run testing
python evaluate_script.py
This script includes the testing process for all systems. The results are stored in each system's output folder.
-
Results statistics
python statistic_faults.py python statistic_hota.py
If Drivence helps your research, please cite the following paper:
@article{gao2026drivence,
title={Drivence: Realistic Driving Sequence Synthesis for Testing Multi-sensor Fusion Perception Systems},
author={Gao, Xinyu and Wang, Zhijie and Feng, Yang and Wang, Chaolan and Zhou, Zhehua and Huang, Yuheng and Ma, Lei and Chen, Zhenyu and Xu, Baowen},
journal={IEEE Transactions on Software Engineering},
year={2026},
publisher={IEEE}
}We will update the code as the paper evolves. Please stay tuned!