Textured Object Detection in Stereo Images using Visual Features

A robot can be easily modified to gather depth data alongside optical images. When combined with an effective object detection framework, this allows the robot to not only identify potentially interesting objects but locate their position and orientation within the world. This information is useful for obstacle avoidance and navigation, locating landmarks for positioning, and for locating a target object in the world. If the detection is accurate enough, the robot can also use the information to locate a grasping point. The goal of this project is to produce a trainable object detector suitable for this setting. During training, visual features from highly textured areas of the image are extracted and stored along with their world-space position to form a stored “view” of the object. When presented with a novel scene the detector will try to match highly textured points with the database of stored object descriptions to find consistent formations of features that would indicate the presence of a known object. This project is supervised by Dr. Gary Bradski at Willow Garage. Willow Garage provided the necessary equipment to capture dense stereo images. Upon completion of the project, the resulting system is planned to be integrated into the Robot Operating System and tested on the Personal Robotics platform currently under development at Willow Garage.