1) Infrared sensor.

We integrated an array of IR LEDs on the pin (male) and an array of IR photodiodes on the funnel (female) to detect the IR LEDs. The guiding with this sensors works for a distance 1000mm - 500mm. With this framework we are able to know only the position. However, we can give different frequencies to each funnel-pin pair to differenciate them and identify them.

The pins in this case require power to light up the IR LEDs, hence, the docking station requires to be powered.

In the video, if only the right funnel detects IR, it keeps moving to the right until the left funnel detets its respective IR LEDs. When both funnels detect the IR LED, this means that the roboat is centered for latching.

2) Infrared camera.

In this configuration the IR elements are not integrated in the latching elements (pin neither funnel). Instead an onboard IR camera tries to find and external IR LEDs tag. The camera is a IR stereo camera with low field of view, meaning it can only detect IR upto 500mm. This configuration requires that the onboard computer process the pose estimation, instead of the previous version with IR LEDs and IR sensor, where the processing is performed by the microcontroller.

Estimating the camera pose from a set of 3D-to-2D point correspondences is known as Perspective from n Points (PnP) (or resection). The minimal case involves three 3D-to-2D correspondences. This is called Perspective from 3 Points (P3P) and returns four solutions.

Robustness can be increased if the LEDs are visible from as many view points as possible. Since we are using infrared LEDs whose wavelength matches the infrared-pass filter in the camera, they appear very bright in the image compared to their environment. Thus, a thresholding function is sufficient to detect the LEDs. With at least four LEDs on the target object and the corresponding detections in the camera image, we can compute the 6 DOF pose of the target object with respect to the camera. Our system can also handle more than four LEDs on the target object.

3) Camera - Tag.

Fiducial markers are commonly use in computer vision (CV) and augmented reality (AR) aplications for detection and identification. In robotic applications, fiducial markers have been of crucial importance for obtaining an accurate pose estimation of the marker. Since, estimating the tag’s 6 Degrees of Freedom (DOF) can lead a robotic arm to grab an item, or guide an autonomous robotic boat to a docking station or to latch another robot, see Figure 1.

There are different families of markers, with circular and squared shapes, and color base [4]. Circular tags such as Intersense [11] and Rune tags [3] provide accurate pose estimation in short distances with a high computational cost. While, squared fiducial tags such as ARTags [6], ARToolkit [10], ArUco [7], AprilTags [12] and AprilTag2 [14] have a low computational cost and can be detected from a further distance.

In robotics, the AprilTags framework has been prefered for having a lower false positive rate and higher detection rates in challenging viewing angles at further distances. However, AprilTags marker systems rely on RGB image space for detection and pose estimation, which is susceptible to the perspective ambiguity noise from sub-pixel detections in their corner locations. Resulting in rotational errors, making the pose estimation challenging without additional information [9].

Who uses AprilTags

Boston dynamics, creates humanoid robots for real case scenarios. MIT in multiple laboratories.

Among squared fiducial marker systems, AprilTags [12] [14] perform the best when detecting smaller markers, in high angle inclination and in different lighting levels [5]. This is the reason of their popularity in the robotics community.

The figure shows the AprilTags framework. If the camera is positioned in front of the tag, the lateral distance dy and angle are zero. While the longitudinal distance between the camera and tag is dx in the X-axis. If the camera changes its position with the same orientation, then the lateral distance dy reflects its change in position on the Y-axis, dx is the same longitudinal distance and angle = 0. If the camera in that position changes its orientation to face the tag, then the lateral distance dy = 0 and the angle reading becomes the true angle between the camera and tag.