Using computer vision data to guide a robot’s movement is fundamentally interesting to a roboticist. Two approaches exist in using the camera input to guide the action of a robot: either separately developed the localization, tracking, and planning pipelines, or coupling computer vision data in the servo loop.

In this thesis we will focus on the latter approach, directly using the features from the detection algorithm to guide the control of a dual axial gimbal, and then extended to the complete platform. Related estimation, frequency analysis, and localization are also explored for better visual servo control performance.

In this thesis, we explored the path to build up the complete visual servo system. The challenges and the solutions are listed in a chronical order, where a...[ Read more ]

Using computer vision data to guide a robot’s movement is fundamentally interesting to a roboticist. Two approaches exist in using the camera input to guide the action of a robot: either separately developed the localization, tracking, and planning pipelines, or coupling computer vision data in the servo loop.

In this thesis we will focus on the latter approach, directly using the features from the detection algorithm to guide the control of a dual axial gimbal, and then extended to the complete platform. Related estimation, frequency analysis, and localization are also explored for better visual servo control performance.

In this thesis, we explored the path to build up the complete visual servo system. The challenges and the solutions are listed in a chronical order, where a one-degree-of-freedom, two-degree-of-freedom, and four-degree-of-freedom schemes are introduced in order. Two archetypal visual servo schemes, position-based visual servo control (PBVS) and image-based visual servo control (IBVS) are tested and compared. The basic methods are introduced upfront, and the advanced topics as the velocity estimator, frequency domain analysis, and the other fine-tune methods are introduced in the later chapters.

We show that image-based visual servo control performs much robust under a weak visual frontend. Multiple fine-tune methods are used, and all algorithm are running onboard in real-time in an indoor environment.