ROS Tensorflow objects detector

A ROS package to run any Tensorflow model for object detection or image segmentation using a camera video feed.

Tested on a Drive PX2 running a SSD-Mobilenet at 19FPS. It takes about 3mins to open a session, so be patient when running this node.

Maintainers

• David Torres: DavidTorresOcana

Requirements

• ROS Kinetic
• Python 2
  • cv2
  • Tensorflow for GPU
  • cv-bridge $ sudo apt-get install ros-kinetic-cv-bridge ros-kinetic-cv-camera
• CUDA and CuDNN

Package description

Node

Main node is detector_node.py

• Publish (default, user selectable): /image_objects_detect (sensor_msgs/Image)
• Subscribe (default, user selectable): /cv_camera/image_raw (sensor_msgs/Image)

Tensorflow model

Model definitions must be placed inside /include/ros_object_detector

As example, the model used is a SSD-Inception detecting potholes

• SSD-Inception and other Object detection models

Models' files:

• Graph: Graph must be saved as a frozen graph .pb file in /include/ros_potholes_detector/name-of-your-model/frozen_inference_graph.pb
• Labels definition: .pbtxt fiel defining objects being detected. Must be placed in /include/ros_potholes_detector/name-of-your-model/object-detection.pbtxt

Parameters

• ~tf_model/pb_path (string: default "pothole_detector/pothole-graph/ssd-inception-30000-27-03-18/frozen_inference_graph.pb") model's checkpoint .pb file path relative to zinclude/ros_potholes_detector
• ~tf_model/labels_definition (string: default "pothole_detector/object-detection.pbtxt") labels' .pbtxt file path relative to /include/ros_potholes_detector
• ~tf_model/num_objects (int: default 1) Number of objects: Number of objects defined in $labels_definition
• ~camera_topic (string: default /cv_camera/image_raw) Subscribed camera topic input to the model: ROS image topic of message type sensor_msgs.msg/Image
• ~out_img_topic (string: default /image_objects_detect) Output image topic name: message type sensor_msgs.msg/Image

Usage

There are three options to use this package.

use a camera:

$ rosrun cv_camera cv_camera_node

1. Using default model and topics

Usig default parameters and model definition:

$ rosrun ros_object_detector detector_node.py

2. Using your model and topics: Use configuration file

If you want to configure the node to set different parameters, you can use roslaunch and /config/default.yaml file.

• Edit /config/default.yaml to set your parameters: Your mdoel's definition, topics, etc
• Use roslaunch:

$ roslaunch ros_object_detector ros_object_detector.launch

or, to launch also the camera node

$ roslaunch ros_object_detector ros_object_detector_all.launch

3. Using your model and topics: on command line

You can also set any of the parameters when calling the node, by using any of the arguments below.

$ rosrun ros_object_detector detector_node.py _tf_model/pb_path:=path-to-your-models-pb.pb _tf_model/labels_definition:=path-to-your-labels.pbtxt _tf_model/num_objects:=your-models-num-objects _camera_topic:=input-camera-topic _out_img_topic:=output-image-topic

E.g.: To use a different camera topic:

$ rosrun ros_object_detector detector_node.py _camera_topic:=input-camera-topic

Note that, otherwise specified, node will use default parameters.

TODOs

• Include other Tensorflow models like YOLOv2
• Output objects detected in a different Topic/format