EMP

EMP is an edge-assisted multi-vehicle perception system for connected and autonomous vehicles (CAVs), designed for efficient and scalable sensor data sharing under potentially fluctuating network conditions to enhance the CAVs' local perception. The code is used in our MobiCom '21 paper, EMP: Edge-assisted Multi-vehicle Perception.

Setup EMP

Setup PointPillars (please first setup the Python environment following this README) and Draco.

cd <src_root>/src/main/cpp/
./setup_pointpillars.sh  # Download PointPillars source code
./build_pointpillars.sh  # Build PointPillars library

cd <src_root>/src/main/cpp
./setup_draco.sh  # Download Draco source code
./build_draco.sh  # Build Draco library

Build the JNI native library.

cd <src_root>/
./build_native_lib.sh

You can directly run the system using IDE like IntelliJ IDEA or generate .jar file for running in a terminal.

./gradlew shadowjar

Run EMP

Edge

First start an edge instance. Change the algorithm index (1-4) to run different partitioning algorithm:

1. No partitioning.
2. Naive partitioning based on the Voronoi Diagram.
3. Bandwidth-aware partitioning based on the Power Diagram.
4. REAP partitioning.

java -Djava.library.path=build/src/main/cpp \
   -Dlog4j.configurationFile=src/main/resources/log4j2-config.xml \
   -cp build/libs/emp-1.0.jar org.emp.edge.EdgeServer \
   -p [server port] -t [number of threads] -a [algorithm index] -c [number of clients] (-s [save path])

Vehicle

Then start vehicle instances in different terminals/machines. The vehicle type corresponds to the algorithm index (1: VehicleFull, 2: VehicleNaive, 3: VehicleBW, 4: VehicleReap). In the dataset object-0227, the available vehicle IDs are 1 (ego-vehicle), 4354, 4866, 5378, 21506, 22786.

java -Djava.library.path=build/src/main/cpp \
   -Dlog4j.configurationFile=src/main/resources/log4j2-config.xml \
   -cp build/libs/emp-1.0.jar org.emp.vehicle.[vehicle type] \
   -i [server ip] -p [server port] -c [client port] \
   -v [vehicle ID] -d [data path] -r [frame rate]

Perception

We modified PointPillars and use it as the perception module in EMP evaluation.

Data

We modified DeepGTAV - PreSIL to create our synthetic multi-vehicle dataset from GTA V. The dataset contains vehicle sensor data (camera images, LiDAR point clouds, position/direction information) and labels.

Data type	Description	Format
image_2	Ego vehicle's front camera data
velodyne	Ego vehicle's LiDAR data	x, y, z, object ID (if the point is in a object such as vehicle or pedestrian)
oxts	Ego vehicle's position/direction information in the KITTI format	latitude, longitude, altitude (KITTI world position), roll, pitch, heading
label_2	Ground truth for surrounding objects	type, truncated, occluded, alpha, 2d bbox (left top, right bottom), dimensions (hwl), 3d bbox (xyz, camera coordinates), rotation_y (camera coordinates)
label_2_aug	Augmented labels for surrounding objects	type, truncated, occluded, alpha, 2d bbox (left top, right bottom), dimensions (hwl), 3d bbox (xyz, camera coordinates), rotation_y (camera coordinates), object ID, points hit in 2D, points hit in 3D, speed, roll, pitch, object model, ID of vehicle the pedestrian is in
ego_object	Augmented label for the ego vehicle (some values set to constant)	type, -1, -1, -1, -1 -1 -1 -1, dimensions (hwl), 0 0 0, -1, object ID, -1, -1, speed, -1, -1, object model, 0

Reference

Please cite our work if it helps your research:

Xumiao Zhang, Anlan Zhang, Jiachen Sun, Xiao Zhu, Y. Ethan Guo, Feng Qian, and Z. Morley Mao. 2021. EMP: edge-assisted multi-vehicle perception. Proceedings of the 27th Annual International Conference on Mobile Computing and Networking. Association for Computing Machinery, New York, NY, USA, 545–558. DOI:https://doi.org/10.1145/3447993.3483242

Contact

Please feel free to reach out ([email protected]) for any questions about how to use the code or access to our dataset.