PointNet Object Detection for ROS2

Train PointNet on KITTI dataset and use it for real-time object detection in ROS2.

Features

• ✅ Train PointNet on KITTI dataset
• ✅ Detect: Car, Bus, Bicycle, Pedestrian
• ✅ ROS2 integration for real-time inference
• ✅ 3D bounding box visualization in RViz

Quick Start

1. Train on KITTI Dataset

python train.py \
    --data_dir ~/datasets/kitti/kitti_dataset \
    --epochs 100 \
    --batch_size 32 \
    --device cuda

2. Use in ROS2

# Build ROS2 package (if needed)
colcon build

# Run detector node
ros2 run pointnet_detector pointnet_detector_node \
    --ros-args \
    -p model_path:=checkpoints/best.pth \
    -p input_topic:=/velodyne_points \
    -p confidence_threshold:=0.5

3. Visualize in RViz

• Subscribe to /pointnet/markers topic
• See 3D bounding boxes for detected objects

Training

python train.py \
    --data_dir ~/datasets/kitti/kitti_dataset \
    --epochs 100 \
    --batch_size 32 \
    --num_points 1024 \
    --device cuda \
    --output_dir checkpoints

ROS2 Node Parameters

• model_path: Path to trained model (default: checkpoints/best.pth)
• input_topic: Point cloud input topic (default: /velodyne_points)
• marker_topic: Output marker topic (default: /pointnet/markers)
• confidence_threshold: Minimum confidence (default: 0.5)
• device: cuda or cpu (default: auto-detect)

Class Mapping

KITTI → Our Classes:

• Car, Van → Car
• Bus, Truck, Tram → Bus
• Cyclist → Bicycle
• Pedestrian, Person_sitting → Pedestrian