Object recognition using local invariant features for robotic applications: A survey

• A complete analysis of object recognition methods based on local invariant features from a robotics perspective is given.
• A brief description of the main approaches reported in the literature is included.
• Methods are analyzed by considering the main requirements of robotics applications.
• Best performing object recognition systems are built using ORB–ORB and DoG–SIFT keypoint-descriptor combinations.

The main goal of this survey is to present a complete analysis of object recognition methods based on local invariant features from a robotics perspective; a summary which can be used by developers of robot vision applications in the selection and development of object recognition systems. The survey includes a brief description of the main approaches reported in the literature, with more specific analyses of local interest point computation methods, local descriptor computation and matching methods, and geometric verification methods. Different methods are analyzed by considering the main requirements of robotics applications, such as real-time operation with limited on-board computational resources, and constrained observational conditions derived from the robot geometry (e.g. limited camera resolution). In addition, various object recognition systems are evaluated in a service-robot domestic environment, where the final task to be performed by a service robot is the manipulation of objects. It can be concluded from the results reported that (i) the most suitable keypoint detectors are ORB, BRISK, Fast Hessian, and DoG, (ii) the most suitable descriptors are ORB, BRISK, SIFT, and SURF, (iii) the final performance of object recognition systems using local invariant features under real-world conditions depends strongly on the geometric verification methods being used, and (iv) the best performing object recognition systems are built using ORB–ORB and DoG–SIFT keypoint–descriptor combinations. ORB–ORB based systems are faster, while DoG–SIFT are more robust to real-world conditions.