Deep Learning Approach For Object Tracking Of RoboEye

Abstract

RoboEye is a spherical 3RRR parallel robot which has been developed for its high precision. It can provide high speeds, so can be used for fast tracking tasks. To this end, in this paper proper deep learning approaches are combined with classical control methods. Deep learning algorithms are employed to detect an object of interest among various ones in a monocular image, and then obtain an estimatation of the distance to the camera. So, simultaneous depth estimation, and object detection with a monocular camera for real time implementation is proposed here. For fast calculations, also to overcome manufacturing uncertainties, inverse kinematic equations are computed by a multi-layer perceptron (MLP) network based on real data. Finally, a classical PID controller can perform a fast tracking of the object.