L3: Visual Object Detection and Classification¶

YOLO Detector Demo - Quick Start Guide

This guide walks you through setting up and running the YOLO object detection demo on the KITTI dataset.

Prerequisites¶

Ubuntu 22.04 or 24.04
ROS 2 Jazzy installed
Python 3.10 or newer
~10 GB free disk space

Step 1: Download the KITTI Rosbag from Google Drive¶

Option A: Using Web Browser

Go to the shared Google Drive link (provided by the instructor)
Download kitti_training.bag (this will be a folder/directory)
Move it to your desired location, e.g., ~/datasets/

Option B: Using gdown (Command Line)

# Install gdown if not already installed
pip install gdown

# Download the bag (replace FILE_ID with actual Google Drive file ID)
gdown --folder https://drive.google.com/drive/folders/FILE_ID

# Or if you have a direct link
gdown https://drive.google.com/uc?id=FILE_ID -O ~/datasets/kitti_training.bag --folder

Option C: Create Your Own Bag from KITTI Dataset

# Download KITTI object detection dataset from:
# http://www.cvlibs.net/datasets/kitti/eval_object.php?obj_benchmark=2d

# Create bag using the provided script
python3 create_kitti_bag_jazzy.py \
    --dataset /path/to/KITTI/training \
    --output ~/datasets/kitti_training.bag \
    --frequency 10.0

Verify the Bag

# Check bag info
ros2 bag info ~/datasets/kitti_training.bag

# Expected output:
# - Duration: ~748 seconds
# - Messages: 7481
# - Topics: /kitti/camera/image

Step 2: Clone and Setup the Package¶

# Create workspace
mkdir -p ~/yolo_ws/src
cd ~/yolo_ws/src

# Clone or copy the yolo_detector package
cp -r /path/to/yolo_detector .

# Or clone from repository
# git clone <repository_url> yolo_detector

Step 3: Install ROS Dependencies¶

cd ~/yolo_ws
sudo rosdep init  # Only if not done before
rosdep update
rosdep install --from-paths src --ignore-src -r -y

This installs: - ros-jazzy-cv-bridge - ros-jazzy-vision-msgs - ros-jazzy-sensor-msgs - ros-jazzy-std-msgs

Step 4: Install Python Dependencies¶

Fix NumPy Compatibility Issue

pip uninstall numpy -y
pip install "numpy<2"
python3 -c "import numpy; print(numpy.__version__)"

Install YOLO and Other Dependencies

cd ~/yolo_ws/src/yolo_detector
pip install -r requirements.txt

Verify Installation

python3 -c "from cv_bridge import CvBridge; print('✓ cv_bridge works')"
python3 -c "from ultralytics import YOLO; print('✓ ultralytics works')"
python3 -c "import cv2; print('✓ opencv works')"

Step 5: Download YOLO Weights¶

Note: They are already in the package’s config folder.

python3 -c "from ultralytics import YOLO; YOLO('yolov8n.pt')"
python3 -c "from ultralytics import YOLO; YOLO('yolov8s.pt')"
python3 -c "from ultralytics import YOLO; YOLO('yolov8m.pt')"

Available Models

Model	Size	Speed	Accuracy	Use Case
yolov8n.pt yolov8s.pt yolov8m.pt yolov8l.pt yolov8x.pt	6 MB 22 MB 52 MB 87 MB 136 MB	Fastest Fast Medium Slow Slowest	Good Better Excellent Great Best	Real-time, embedded Balanced Accuracy-focused Best accuracy Maximum accuracy

Step 6: Configure Parameters (Optional)¶

Edit parameter file:

nano ~/yolo_ws/src/yolo_detector/config/yolo_params.yaml

Example:

yolo_detector:
  ros__parameters:
    model_path: "yolov8n.pt"
    conf_threshold: 0.5
    iou_threshold: 0.45
    image_topic: "/kitti/camera/image"
    max_det: 300
    device: ""

Pre-configured Files:

yolo_params.yaml — Default
yolo_params_sensitive.yaml — More detections
yolo_params_conservative.yaml — Fewer detections

Step 7: Build the Package¶

cd ~/yolo_ws
colcon build --packages-select yolo_detector
source install/setup.bash
echo "source ~/yolo_ws/install/setup.bash" >> ~/.bashrc

Verify build:

ros2 pkg list | grep yolo_detector
ros2 pkg executables yolo_detector

Step 8: Launch the Demo¶

Basic Launch

ros2 launch yolo_detector yolo_detector.launch.py \
    bag_file:=~/datasets/kitti_training.bag \
    play_rate:=0.1

Playback Speed Options

play_rate:=1.0 — Normal speed
play_rate:=0.5 — Recommended for viewing
play_rate:=0.1 — Good for analysis

Manual Launch (Separate Components)

# Detector node
ros2 run yolo_detector yolo_detector_node --ros-args --params-file ...
# Bag player
ros2 bag play ~/datasets/kitti_training.bag --loop --rate 0.5
# RViz
rviz2 -d $(ros2 pkg prefix yolo_detector)/share/yolo_detector/config/demo.rviz

Step 9: Visualize with Foxglove Studio¶

Install Foxglove Bridge

sudo apt install ros-jazzy-foxglove-bridge
ros2 launch foxglove_bridge foxglove_bridge_launch.xml

Connect to ROS 2

Open Foxglove Studio (Desktop or Web)
Connect to ws://localhost:8765

Load Demo Layout

Import demo.json from config/

Panels Overview

Raw Image → /kitti/camera/image
Detections Image → /detections/image
Detection Count → /detections/count
Inference Time → /detections/inference_time
FPS → /detections/fps
Stats (JSON) → /detections/stats

Expected Output¶

In RViz:

Two image panels showing raw and detected images.

In Foxglove:

Two image panels showing raw and detected images.
Real-time plots (detections, FPS, inference time).

In Terminal:

[INFO] [yolo_detector]: Frame 30: 5 objects | Inference: 45.2ms | Avg: 47.3ms (21.1 FPS)

Monitoring and Debugging¶

ros2 topic list
ros2 topic hz /kitti/camera/image
ros2 topic hz /detections
ros2 param list /yolo_detector
ros2 param get /yolo_detector conf_threshold
ros2 param set /yolo_detector conf_threshold 0.3

Troubleshooting¶

NumPy Version Error

pip uninstall numpy -y
pip install "numpy<2"

Package Not Found

colcon build --packages-select yolo_detector
source install/setup.bash

YOLO Model Not Found

python3 -c "from ultralytics import YOLO; YOLO('yolov8n.pt')"

Stopping the Demo¶

Ctrl+C
pkill -9 ros2
pkill -9 rviz2