local_mapping_lidar

Local Mapping algorithm for LiDAR sensor and monocular camera.

The algorithm is as follows:

• The ground plane is removed from the LiDAR point cloud using RANSAC.
• The outlier points (not on the ground plane) are clustered using DBSCAN.
• On the camera image, a bounding box is cropped for each detected cone of the point cloud,, by knowing the LiDAR resolution and the camera model.
• The cone is classified by its class (or as a "not cone") by a simple convolutional neural network.
• For each cone classified, associate its class to the previous detected position.

Inputs

• /pointcloud (sensor_msgs/PointCloud2)
• /image (sensor_msgs/Image)
• /camera_info (sensor_msgs/CameraInfo)

Outputs

• /cones (lart_msgs/ConeArray)

Required transforms

• base_link -> lidar_link
• base_link -> camera_link

Dependencies

• ROS Humble
• CMake
• C++ compiler
• Open3D
• LibTorch
• lart_msgs