9. Object Detection

9.1. System Description

This ti_vision_cnn node is a versatile deep-learning (DL) inference ROS node that is optimized on DL cores and hardware accelerator of TI Processors. The ti_vision_cnn node supports compute-intensive DL inference operations, including 2D object detection and semantic segmentation. Figure 1 shows the high-level block diagram of the applications around the ti_vision_cnn node, which consists of multiple processing blocks that are deployed on hardware accelerators and DSP processors for pre-processing and post-processing in an optimized manner.

Figure 1. Object detection demo: block diagram

For details of block diagram and parameters of ti_vision_cnn, please refer to README.md.

Note

cam_id and subdev_id for cameras: You can check the cam_id and subdev_id for the camera attached to the SK board by running /opt/edgeai-gst-apps/scripts/setup_cameras.sh. Accordingly please update the parameters or pass as launch arguments.

Note

IMX390 Camera: Before running, please refer to gscam2/README_TI.md for generating required LUT files for tiovxldc.

9.2. Run the Application in ROS 2

[SK] To launch the object detection demo with a ZED stereo camera, run the following command inside the Docker container on the target:

ros2 launch ti_vision_cnn zed_objdet_cnn_launch.py cam_id:=x zed_sn:=SNxxxxx

To process the image stream from a USB mono camera:

roslaunch ti_vision_cnn gscam_objdet_cnn.launch cam_id:=x

For IMX219 camera as input,

ros2 launch ti_vision_cnn gscam_objdet_cnn_imx219_launch.py cam_id:=x subdev_id=y

For IMX390 camera as input, depending on choice of resolution, run one from the following commands.

ros2 launch ti_vision_cnn gscam_objdet_cnn_imx390_launch.py cam_id:=x subdev_id=y width:=1920 height:=1080
ros2 launch ti_vision_cnn gscam_objdet_cnn_imx390_launch.py cam_id:=x subdev_id=y width:=1280 height:=720
ros2 launch ti_vision_cnn gscam_objdet_cnn_imx390_launch.py cam_id:=x subdev_id=y width:=960 height:=540

[Visualization on Ubuntu PC] For setting up environment of the remote PC, please follow Docker Setup for ROS 2

Depending on the resolution setting of the capture node, run one from the following:

ros2 launch ti_viz_nodes rviz_objdet_cnn_launch.py width:=1920 height:=1080
ros2 launch ti_viz_nodes rviz_objdet_cnn_launch.py width:=1280 height:=720 $ default for USB camera and IMX219
ros2 launch ti_viz_nodes rviz_objdet_cnn_launch.py width:=960 height:=540