Embedded Motion Control 2013 Group 10

Group Name

Team Picobello

Group Members

Name:	Student id:	Email:
Pepijn Smits	0651350	p.smits.1@student.tue.nl
Sanne Janssen	0657513	s.e.m.janssen@student.tue.nl
Rokus Ottervanger	0650019	r.a.ottervanger@student.tue.nl
Tim Korssen	0649843	t.korssen@student.tue.nl

Introduction

Nowadays, many human tasks have been taken over by robots. Robot motion control and embedded motion control in the future allow us to use robots for many purposes, such as health care. The (humanoid) robots therefore have to be modeled such that they can adapt to any sudden situation.

The goal of this project is to get the real-time concepts in embedded software design operational. This wiki page reviews the design choices that have been made for programming Jazz robot Pico. Pico is programmed to navigate fully autonomously through a maze, without any user intervention.

The wiki contains three different programs. The first program was used for the corridor competition. In this program the basic skills like avoiding collision with the walls, finding corners and turning are implemented.

After the corridor competition, a new design strategy was adopted. The second program consists of a low level code, namely a wall follower. By keeping the right wall at a fixed distance from Pico, a fairly simple, but robust code will guide Pico through the maze.

Besides this wall follower strategy, a high level approach was used. This maze solving program uses wall (line) detection to build a navigation map, which Pico uses to create and follow path lines to solve the maze. To update the position of Pico, the odometry information, local and global maps are used.

The latter two programs use Pico’s camera to detect arrows in the maze that point Pico in the right direction.

Data processing

Pico outputs a lot of sensordata. Most of the data needs to be preprocessed for use the maze-solving algorithm. The odometry data, laserscandata and imagedata are discussed below.

Odometry data

One of the incoming data types is the odometry. The odometry information is based on the angular positions of the robot wheels. These angular positions are converted into a position based on a Cartesian odometry system, which gives the position and orientation of Pico, relative to its starting point. For the navigation software of Pico, only the x,y-position and [math]\displaystyle{ \theta }[/math]-orientation are of interest.

Due to slip of the wheels, the odometry information is never fully accurate. Therefore the odometry is only used as initial guess in the navigating map-based software. For accurate localization, it always needs to be corrected.

This correction is based on the deviation vector obtained from the map updater. This deviation vector gives the (x,y,θ)-shift that is used to fit the local map onto the global map and therefore represent the error in the odometry. Because of the necessary amount of communication between the parts of the software that create the global map and the part that keeps track of the accurate position, these parts are programmed together in one node. This node essentially runs a custom SLAM (Simultaneous Localization and Mapping) algorithm.

Laserscan data

To reduce measurement noise and faulty measurements the laser data is filtered. Noise is reduced by implementing a Gaussian filter, which smoothens each data point over eight other data points.

[math]\displaystyle{ G(x)=\frac{1}{\sqrt{\sigma \pi}}e^{\frac{-x^2}{2 \sigma^2}} }[/math] with [math]\displaystyle{ \sigma = 2.0 }[/math]

Faulty measurements are eliminated in three different ways. First data points which deviate a lot from their neighbors are ignored in the Gaussian filter so that the filter does not close openings. When ignoring data points the Gaussian is normalized based on the used points.

Secondly the first and last 15 data points are ignored, decreasing the angle of view by 7.5 degrees at each side, but preventing the robot to measure itself. At last, only data points are used which are between a certain range, not only to prevent measuring the robot itself, but also to eliminate outliers.

Arrow detection

Binary image

From a full color camera image we want to determine whether there is an arrow and to determine its direction. Because this full color camera image is too complex to process, we transform it into a binary image. Since the arrows used in the competition are red, everything that is red (with a certain margin) is made white and the remaining is set as black.

Canny edge detection

Next edges are detected with a canny edge detection algorithm from openCV. This algorithm uses a Gaussian filter to reduce noise, similar to the one explained above. But it also uses the gradient of this Gaussian filter. At and edge there is a very strong transition from white to black, resulting in an extreme value in the gradient. So by thresholding the gradient, edges can be detected.

Find contours

The binary image from the canny edge detection is used as input in the FindContours function from openCV. This function finds contours by connecting nearby points, giving a vector of contours as output. Each contour consists of a vectors filled with points.

Arrow detection

figure x.x Arrow direction

figure x.x Arrow shape

Finally the image is processed so that we can start detecting arrows. Arrows are detected in two steps. First the ratio of the height and the width of each contour is determined, to see if the contour can be an arrow, see figure 1. The arrow used in the competition has a ratio of 2.9, with a lower and upper bound, this results in: [math]\displaystyle{ 2.4\lt \frac{dx}{dy} \lt 3.8 }[/math]

Next if a contour can be an arrow, the direction is investigated. This is done by determining the ratio of dx_right and dx_left, see figure 2. If the largest width (dy) is shifted towards the right dx_left > dx_right and the other way around. With some lower and upper bound this results into:

Right arrow if:

[math]\displaystyle{ dx_{left} \gt \frac{5}{4} dx_{right} }[/math]

Left arrow if:

[math]\displaystyle{ dx_{left} \lt \frac{4}{5} dx_{right} }[/math]

No arrow if:

[math]\displaystyle{ \frac{4}{5} dx_{right} \lt = dx_{left} \lt = \frac{5}{4} dx_{right} }[/math]

If the largest width is in between the margins, the contour cannot be an arrow and is treated as such. This may not be the most robust way to detect an arrow, but during the tests it seemed to work just fine. A few other methods that can be used or combined to make the arrow detection more robust are circularity, shape matching, matching of moments or finding lines with the Hough transform.

figure x.x Camera image	figure x.x Binary image	figure x.x Canny edge detection
figure x.x Contours	figure x.x Arrow detection

Program 1: Corridor Challenge solver

Edge detection

For the corridor competition, Pico should first be able to detect an outside corner (see picture 1). To detect corners, the laser range data is used.

The robot searches left and right for large differences in wall distance. If the length ratio between two neighboring ranges exceeds a certain threshold, the position (1) of the outside corner is saved. Also the side of the corridor is set to left or right. If the first corner has been found, Pico searches at this side for other corners. This search algorithm starts at the first corner and checks for the shortest laser range distance. This coordinate is saved as second edge position (see picture 2). The coordinates of the edges are sent to the make turn function, which controls Pico’s movement.

If Pico is past the first corner, the first algorithm explained above cannot recognize this first corner anymore. To avoid problems, the second algorithm is activated for both the first and last corner (see picture x.x).

figure x.x Detect first edge

figure x.x Detect first edge

figure x.x Detect first edge