Object detection

Abstract

object_detection is a ROS node that provide an interface to detect object in an image using tensorflow. It takes as input:

• A network file (.pb),
• A label file (.pbtxt),
• An input image topic,
• An output object detection topic.

Each object is represented in the image by:

• A score (float32) in the range [0, 1] that indicate the probability that this detection belongs to the given category,
• A category (string),
• A 2d bounding box in opencv image coordinate system.

Here is a small example for person detection.

Further improvments

We could accelerate the detection pipeline by using a dedicated TPU such as Google Coral that would allow to develop a Kalman Filter to track the object.