FREE ELECTRONIC LIBRARY - Dissertations, online materials

Pages:   || 2 |

«Advanced Mining Technology Center & Dept. of Elect. Eng., Universidad de Chile {pcano,jruizd} Abstract. The main goal of this paper is ...»

-- [ Page 1 ] --

Robust Detection of White Goals

Pablo Cano, Yoshiro Tsutsumi, Constanza Villegas and Javier Ruiz-del-Solar

Advanced Mining Technology Center & Dept. of Elect. Eng., Universidad de Chile


Abstract. The main goal of this paper is to present a simple, but robust

algorithm for detecting white goals in the context of the RoboCup SPL

(Standard Platform League). White goals will be used for the first time in the

SPL competitions in 2015. The main features of the algorithm are a robust search strategy for detecting the goal posts, and the use of the Y channel image, instead of the color segments, for determining and characterizing the goal posts and the horizontal crossbar. This last aspect is crucial for detecting a white goal placed in a white background. The algorithm is validated in the real world (real robot in a SPL field), showing its ability to detect the while goals even when they are observed in a white background.

Keywords: Robot soccer, SPL league, color vision systems, object detection.

1 Introduction As everybody knows in the soccer robotics community, the detection of the field’s objects (goals, lines, ball, players) is essential for playing soccer properly. Given that, the robust detection of objects in dynamic and cluttered environments is a complex task, and considering that legged soccer robots have normally low-computational capabilities, at the beginning of the RoboCup soccer competitions the use of colored objects and beacons was introduced, with the purpose of facilitating the use of color- based vision systems.

Given that the final goal of the RoboCup is to build robots that will be able to compete and defeat humans in real-world fields, the mentioned constraints are relaxed from year to year. For instance, at the beginning of the former Four-Legged league, which after moving from 4-legged robots to humanoid robots is now called Standard Platform League (SPL), 6 colored beacons and 2 colored and solid goals were used.

The 6 colored beacons were first reduced to 4, then to 2, and finally they disappeared.

The colored and solid goals were first transformed into non-solid goals composed by 2 goal posts and 1 horizontal crossbar. Then the color of both non-solid goals was set to yellow, and this year (2015), the goals will be white for the first time. That means that they will be similar to the goals used in most human soccer matches.

The use of white goals is more challenging than the use of yellow ones, because in the image space they can be mixed up with the field lines, but also because in many human indoor environments the walls are white.

In this context, the main goal of this short paper is to present a simple, but robust algorithm for detecting white goals in the context of the SPL. The main features of the algorithm are a robust search strategy for detecting the goal posts, and the use of the Y channel image, instead of the color segments, for determining and characterizing the goal posts and the horizontal crossbar. This last aspect is crucial for detecting a white goal placed in a white background.

The proposed algorithm is validated in a real SPL field. The article is organized as follows: Section 2 describes relevant related work. Section 3 describes the proposed algorithm, and Section 4 the experimental results. Finally, conclusions are drawn in Section 5.

2 Related Work In the RoboCup Soccer community the visual detection of field objects is a well know problem, which has been tackled by most of the teams. Some seminal works are the following. In 1999 Bandlow [1] developed a vision system based on color classification for object recognition and localization in a RoboCup scenario. Using the previous knowledge of the objects and their respective colors, the image is segmented in one of those possible colors, in order to find object candidates. Then, strictly defined constraints of color and shape are stored in a database and compared with the possible objects. Jamzad [2] introduced a novel color model for shape and object detection. They use “jump points” to take advantage of the perspective information of the image in order to obtain a fast object recognition process. Zagal [3] tries to automatize the object detection problem by using supervised learning approaches.

Loncomilla and Ruiz del Solar [4] describe a complete different object recognition approach based on the use of interest points and descriptors. Hartl [5] adds robustness to the problem of objects detection using color similarities. Other works addressing important aspects such as color invariance, the use of natural light, robustness, and the use of context information are described in [6,7,8,9,10,11,12].

Specifically in the goal detection problem, Canas [13] use means of color based segmentation and geometrical image processing methods to detect goals and determine the positions of the robot according to it.

The here-proposed approach is based on the B-Human code release [14], which is based on [15]. In this approach, a new color segmentation method is proposed, where colors are mapped not only to unambiguous but also to ambiguous color classes. In order to find the goal post, the image is horizontally scanned at the projected horizon height of the robot, in order to find yellow pixels that represent a possible goal post (because until last year all goals were yellow). Then, after finding the edges of the goal post candidates, several filters are applied in order to discard the false positive detections. As mentioned, the proposed algorithm is based in [15], but the goal post is not searched at the projected horizon height of the robot, but at the field boundary. In addition, the goal is characterized using the original Y channel image, and not the segmented one.

3 White Goal Detection Algorithm In order to find a white goal in the image under analysis, the following steps are made: Detection of Posts Bases, Determination of Posts Heights, Determination of Posts Widths, Filtering and Characterization (see Algorithm 1). These steps allow finding any possible post inside the image, to filter out the ones that do not represent a real goal, and to characterize the real ones considering the detected posts.

–  –  –

3.1 Detection of Posts Bases In former approaches developed for solving this problem, a horizontal scan through the image is carried out, and every white-segmented pixel is analyzed as a possible goal-post-base candidate. The scanning needs to be done in only one row of the image and, as long as this row stays below the projected horizon of the robot, any existing goal post in the image should cross the scan line.

However, the challenge is to distinguish the white goal-post-base in a white background. This can be addressed by a proper selection of the scanning line, which does not need to be horizontal anymore. In the proposed algorithm a scan through the field boundary is performed. This boundary is a convex hull of the green pixels that represent the field. As it can be seen in Figure 1, the goal post bases are always below the field boundary line, so this can be used as a scan line that will always cross the base of a goal post. Also, the background of this scan line will be always green, so it is still possible to use the segmented image to find the goal post's base.

The proposed algorithm is shown in Algorithm 2. The function yBoundaryValue(x) returns the row value of the field boundary for a given column x.

The algorithm searches white segments by finding first the transition between a green and a white pixel, and then the transition between a white and a green pixel. Next, it uses the transition coordinates as the start and the end of a white segment (goal post candidate). The center of the base of the goal post candidate is calculated by averaging these two points.

–  –  –

Algorithm 2. Pseudo code of the algorithm that scans the image searching for white segments that correspond to candidates of goal posts.

3.2 Determination of the Posts Heights After detecting a possible goal post's base, a bi-directional vertical scan is performed, in order to find the edges of the goal post. Then, the upper edge of the goal post is scanned horizontally in order to determine if it is a left or a right goal post.

When scanning in the downward direction, no major problem exists, because the pixels inside the field are correctly segmented, then, finding the green base is an easy task. However, when going upwards, the white-background problem shows up, because a lot of non-goal post's pixels could be segmented as white pixels. Therefore, while the scan is performed, the image's real intensity values are considered, because small intensity gradients could indicate the end of the goal post. In Figure 2 various shades of white are measured in the Y, Cb and Cr channels. It can be seen that the intensity gradients are more evident in the Y channel.

Fig. 2. Shades of white measurements in different channels. The blue line on the image indicates the position where the pixels values were evaluated. The graphs indicate the value of each channel of the YCbCr color space.

–  –  –

Algorithm 3 shows the pseudo code of the algorithm that scans the image in any direction using the pixel's intensity information. The function f represents the policy used to detect changes in the Y channel. It could be an absolute value between one pixel and the next one or an absolute value between the current pixel and all the previous ones low pass filtered by a Gaussian kernel. The function moveForward implements different ways of scanning the image, according to the current direction. It only moves inside the image and it handles the noise in the segmentation, using a configurable step size hysteresis. For the upward scan, it handles the possible tilt of the post, which could appear when the robot is in motion, or when the post is watched from a certain perspective. In Figure 3 a tilted post is shown. As it can be observed, the

–  –  –

Fig. 3. Example of an upward scan in a tiled post. The function moveForward() recalculates the width and allows to continue the scan. It also clips the scan inside the image.

Then, using Algorithm 3 it is possible to determine the height of the post. Figure 4 shows an example where a traditional approach, which uses only the color-segmented image, and the proposed one, which uses the color-segmented image and the pixel values in the Y channel, are used for determining the goal post edges. It can be seen that only the proposed approach carried out this task properly, even when the background is white.

Fig. 4. Comparison between two different approaches used for finding the goal post's edges. Left, only the color-segmented image is used. Right, the color-segmented image and the pixel's intensity information are used.

3.3 Determination of the Posts Widths After finding the upper and lower edges of the goal post, a line is drawn between these edges. Then, the goal post's width is analyzed in a discrete number of points inside the plotted line. For that purpose, in every chosen point of the initial line, a bidirectional horizontal scan is made. In this step, the same algorithm shown in Algorithm 3 is used.

Figure 5 shows the difference between scanning the image when using and not using the pixel's intensity information to find the goal post's width. Even though the algorithm did not work perfectly, the detection can be done, thanks to the filtering algorithm described in the next section. Also, the detection of the goal post's width in a white background environment cannot be done if the pixel's intensity information is not used.

Fig. 5. Goal post's width detection example. Left, only the color-segmented image is used. Right, the color-segmented image and the pixel's intensity information are used. The green and red lines represent good and bad goal-post width detections, respectively, according to the expected width of a goal post in the image. The green lines are obtained during the filtering step.

3.4 Filtering

Finally, all the possible posts are filtered out, in order to eliminate false positive detections. The principal filter uses the information of the lower edge and the width of the possible goal post to do the classification. First, the distance of the goal post to the robot is calculated using its lower edge position on the image, and the position of the projected horizon on the image. This allows calculating a scaled orthographic projection of the lower edge of the post that gives an estimated position of the post foot in local coordinates. Then, since the actual goal post proportions are known, the calculated goal post's width value can be compared with the expected one, which can be calculated using the goal post’s distance. In Figure 5, the width values that correspond to the expected width values are shown in green, while the wrong ones are shown in red. Thus, using the proportions between the correct and incorrect width values, the false positives are filtered out.

3.5 Characterization The last step consists in characterizing a goal using all the post detections made before. To do this, a simple heuristic is applied, given that in one frame only one goal can be seen, so it is possible to watch only two posts at the same time. The distance between these two posts is also used to discard false positive detections, because the 7 aspect ratio of a complete goal is known. Then, using the distance information of each post, the goal distance and angle to the robot is calculated. This steps are skipped if only one post is detected, and if more than two post are detected, all the detections are discarded.

4 Results The described algorithm was tested using two real videos collected by the robot inside a SPL field. The first video considered only non-white backgrounds, and the second one white backgrounds. Please note that all the statistics shown below were calculated using the detections made by the robot itself.

The first video contained 1,873 frames, non-white backgrounds, and the goal was visible in 576 of the 1,873 frames. A person moved the robot to different positions on the field. Like so, the robot was able to look the goals and other parts of the field.

Pages:   || 2 |

Similar works:

«Lakes A Zone Design Guide for Buildings 1 LAKES A ZONE DESIGN GUIDE for BUILDINGS September 2002 2 LAKES A ZONE DESIGN GUIDE for BUILDINGS September 2002 Low profile buildings of simple, contemporary form, and mid to dark finishes, can be unobtrusively nestled in to the lower slopes and vegetation of the lakes landscapes. Even with glazed front facades, large overhangs can provide shading to prevent these glinting in the sunlight. Rotorua District Council Private Bag RO 3029 Rotorua tel. (07)...»

«DOI: 10.2478/auom-2014-0037 An. St. Univ. Ovidius Constanta ¸ ¸ Vol. 22(2),2014, 121–127 ON THE SOLVABILITY IN LOCAL EXTENSIONS Akram Lbekkouri Abstract The aim of this paper is the study of the solvability of finite extensions of local fields in the general case that is when the residue field is not necessarily perfect. 1 Perfectness of residue fields in ramification theory In case of a perfect residue field or at least when the residue extension is assumed to be separable (ie....»



«Outdoor Action Outdoor Adventure Trip Personal Equipment List Below is the list of equipment you will need to bring with you for your Outdoor Adventure trip. The following list of equipment is based on thirty years of experience in running trips. You will be quite comfortable in any situation if you bring all the gear on the list. You shouldn’t need to bring other items. Generally, we have summer temperatures in the 80s to 90s ° F. But temperatures can vary considerably cooler — in the 60s...»

«Archaeology and cultural belonging in contemporary Syria: the value of archaeology to religious minorities Emma Loosley Abstract This paper has evolved out of a series of community projects that I have initiated in Syria since 1997. My hypothesis is that a knowledge of the archaeological remains and religious traditions of a region bring local people more in touch with their cultural roots and encourage a sense of belonging. This is particularly the case when a religious minority feels...»

«31 2002 TABOO London run: The Venue. January 29th (87 performances) Music: Kevan Frost, Richie Stevens & John Themis Book & Lyrics: Mark Davies & Boy George Director: Christopher Renshaw Choreographer: Les Child Musical Director: James McKeon Cast: Euan Morton (Boy George), Matt Lucas (Leigh Bowery), Luke Evans (Billy), Dianne Pilkington (Kim), Mark McGee (Marilyn), Drew Jaymson (Steve Strange), Paul Baker (Philip Sallon), Gemma Craven (Josie), Mark White (Derek/Petal), Gail Mackinnon (Big...»

«Silo Sense Table of Contents Page Introduction 1 Unloader Suspensions and Cables 2 Leaning Silos 3 Silo Chutes 4 Silo Doors 5 Fill Pipes 6 Hydric Lifting 7 Concrete Deterioration 8 Silo Roofs 9 Silo Hoops 10 Feed Distribution 11 Silo Insurance 12 Silo Ladders and Cages 13 Unloader Service 14 Shell Corn in Conventional Silos 15 Overfilling Your Silo 16 Silo Footing Concerns 17 Silo Footings 18 Introduction Like soldiers standing at attention, upright silos stand all over the country preserving...»

«EuroScooter – OWNER’S MANUAL 1 TABLE OF CONTENTS Conformity Introduction Scope of delivery Purpose of use General safety instructions, warnings and cautions Safety and traffic regulations Vehicle load capacity Maintenance / general information Delivery check Personalized settings Seat Handelbar position Operation Brake Instrument Cluster and lighting Brakes Handelbar Basket Electronic Charging the Battery Battery replacement Cleaning Inspection / Maintenance Lubricant specification Brake...»

«“Sometimes, to change the people, you’ve got to change the people”: When learning is not enough Abstract: “Sometimes, to change the people, you‟ve got to change the people.” A leader‟s comment prompts us to consider whether change that requires a quick shift to new schematic sets, aspirations, activities, goals and capabilities will require radical measures. This paper describes change management models and organisational learning strategies commonly cited in change literature. It...»

«Grassroots Post-Modernism Remaking the soil of cultures Gustavo Esteva & Madhu Suri Prakash 1 'Beyond its definite No to the Global Project, this book takes a stimulating glance at the renewed life of social majorities and offers good reasons for a common hope! GILBERT RIST 'Grassroots Post-modernism is daring in its thesis that the real postmodernists are to be found among the Zapotecos and Rajasthanis of the majority world. It is hard-hitting in its attacks against progressive commonplaces,...»

«Community Development Department 18400 Murdock Circle | Port Charlotte FL 33948 Building Phone: 941.743.1201 | Building Fax: 941.764.4907 Zoning Phone: 941.743.1964 | Zoning Fax: 941.743.1598 BuildingSvcs@CharlotteCountyFL.gov www.CharlotteCountyFL.gov ”To exceed expectations in the delivery of public services” Residential Pool/Spa Fact Sheet Print Form Applicant is responsible for compliance with all applicable Federal, State, and Local codes; some common code articles have been included...»

<<  HOME   |    CONTACTS
2016 www.dissertation.xlibx.info - Dissertations, online materials

Materials of this site are available for review, all rights belong to their respective owners.
If you do not agree with the fact that your material is placed on this site, please, email us, we will within 1-2 business days delete him.