Perception: Multi-Object Detection and Tracking

1. Multi-Object Detection and Tracking for Autonomous Driving / Robot Perception
2. LiDAR 3D object detection using PCL
3. Camera 2D object detection using YOLOv3 model
4. Tracking and sensor Fusion - ongoing
5. Docker and GitHub actions are used for CI/CD
6. GTest is used for executing tests.
7. ROS is used for visualization

tracking.mp4

Dependencies

• Eigen3
• OpenCV
• PCL 1.8
• ROS Noetic
• GTest
• Boost (with components: filesystem, thread, system, program_options)
• Download KITTI tracking dataset, convert.bin pointcloud files to .pcd file

Usage

Detection and Tracking

catkin_make
source devel/setup.bash
rosrun motl main_node DATA_PATH

Testing

rosrun motl test_main_node 

Docker

docker build -t mot .
docker run --network="host" -v DATA_PATH:/dataset -it mot
rosrun motl main_node /dataset

run roscore inside container first and replace dataset valume path in the dockerfile.

LiDAR Object Detection:

Uses classical Poincloud Processing algorithms in PCL to detect 3D obstacles.

1. Filtering - Downsampling, Outlier removal, ROI cropping
2. Segmentation of ground plane
3. Clustering of obstacles
4. Bounding box detection around obstacles
5. Visualization of detected obstacles and ground plane

Camera Object Detection:

Uses YOLOv3 object detection model to detect 2D obstacles.

1. Image Preprocessing
2. YOLO Object Detection
3. Non-max suppression
4. Visualization of detected obstacles

Mulit-object Tracking:

1. Extended Kalman Filter for Tracking and Fusion
2. Hungarian Algorithm for data association
3. Track Management: SORT algorithm