[melodic-devel] Add voxel grid filter and support for processing points with masks

[Alpaca-zip]

PR Type

• Feature
• Bug fix
• Refactor
• Documentation
• Other

Overview

• Addition of voxel grid filter in related to tracker_with_cloud_node not returnn 3D bounding box on topic /detection3d_result  #34.
• Add support for processing points with masks.
• Fix to fuse models to reduce yolo start-up time.
• Add parameter to enable gpu device to be specified in related to issue Running on GPU #33.
• Changed the use of the classes parameter.
• Changed the type of messages that tracker_node publishes.
• Add PR template.

Detail

• The type of messages published by tracker_node has been changed as follows.

  • ~result_topic

    std_msgs/Header header
    vision_msgs/Detection2DArray detections
    sensor_msgs/Image[] masks
    

  • ~result_image_topic(sensor_msgs/Image)

• If you are using segmentation(e.g. yolov8m-seg.pt), tracker_with_cloud_node uses the mask image in project cloud.

  tracker_node	tracker_with_cloud_node

• If you want to run yolo with a specific class id, modify the launch file as follows.

  <arg name="classes" default="[0, 1]"/> <!-- person, bicycle -->

• If you want to specify the device to be used by yolo (e.g. cpu or cuda:0), modify the launch file as follows.

  <arg name="device" default="cpu"/> <!-- cpu -->

  <arg name="device" default="0"/> <!-- cuda:0 -->

Test

• I have tested this change with KITTI data set.

Attention

• If the 3D bbox is not displayed correctly in tracker_with_cloud_node, please consider using a lighter yolo model or increasing the voxel_leaf_size.
• For other changes, please refer to the README or check the commit.