Fast Soccer Ball Detection using Deep Learning

Transcription

1 Fast Soccer Ball Detection using Deep Learning Spring 2017 Slides by Aref Moqadam Mehr 1

2 Problem Statement RoboCup Soccer League NAO HSL Field is no longer Color-Coded Unspecified Patterns (in HSL) NAO white robot in addition to all the problems 2

3 Intuition Image Image Acquisition Color Analysis Field Boundary Circle Detection Verification Image Acquisition Color Analyze Field Boundary Circle Detection Verification Hopefully the proper Result! 3

4 Color Analyze Field Boundary Detection 4

5 Field Boundary Field Boundary Detection Ball is always inside the field Outside of the field is usually crowded with patterns. 5

6 Field-Green Color When the robot is inside the field usually it see mostly green (When the image is cropped below the horizon). Horizon Thomas Reinhardt 6

7 Horizon In graphical perspective, a vanishing point is a point in the image plane where the projections of a set of parallel lines in space intersect. -Wikipedia Horizon is the projection of the infinity (or a very distanced point) Photo by: 7

8 In our work The highest pixel for field boundary in image is when robot is at one corner and observing the diagonal corner of the field. Infinity = ~10meters Project a point in 10m distance = horizon Photo by: 8

9 Cr Peak Again Field-Green Color p = peak(histogram) isgreen(p) = P p < δ

10 Field Boundary Run Ups Results Andrew's monotone Results 10

11 Edge Detection An essential Step through ball detection 11

12 Edge Detection 12

13 Edge Detection 13

14 Edge Detection 14

15 Here is the magic Distributed of Random Points Very Efficient 15

16 Circle Detection Main step of the algorithm 16

17 First Things First! Circle Detection 17

18 First Things First! Circle Detection CHT RHT FRHT 18

19 Circle Hough Transform 19

20 Circle Hough Transform 20

21 Circle Hough Transform 21

22 Circle Hough Transform 22

23 Circle Hough Transform 23

24 Circle Hough Transform But What is R?!! 24

25 Circle Hough Transform Since distance effect on the R value But What is R?!! 25

26 Multi Radius Hough Detection Very inefficient both in terms of processing power and memory Hough Space (R 1 ) 26

27 Randomized Hough Transform How it works? 27

28 Randomized Hough Transform How it works? 28

29 Randomized Hough Transform How it works? 29

30 Randomized Hough Transform How it works? i.e. Overlap: 10% i.e. Overlap: 98% 30

31 Randomized Hough Transform Much Faster But does not have enough accurate Require lots of iteration to guarantee an acceptable result. 31

32 Randomized Hough Transform Less Probable (i.e. 0.01%) More Likely to happen (i.e %) 32

33 Fast Randomized Hough Transform To solve the accuracy. Hence, reduce the maximum iteration required Thus, become even faster! 33

34 Fast Randomized Hough Transform 34

35 Fast Randomized Hough Transform 35

36 Fast Randomized Hough Transform 36

37 Fast Randomized Hough Transform 37

38 Fast Randomized Hough Transform 38

39 Fast Randomized Hough Transform 39

40 Fast Randomized Hough Transform 40

41 Fast Randomized Hough Transform There is an extracted circle for each iteration Here in our project: Filter by size Filter by green and non-green pixel percentage inside the circle 41

42 Validation Final Step 42

43 Finally Filters Size White Pixels Maximum Non-Green Pixels Limit Projected Size Pattern 43

44 Finally Filters Size White Pixels Maximum Non-Green Pixels Limit Projected Size Pattern 44

45 Finally Filters Size White Pixels Maximum Non-Green Pixels Limit Projected Size Pattern 45

46 Finally Filters Size White Pixels Maximum Non-Green Pixels Limit Projected Size Pattern 46

47 Pattern Recognition How to precisely find the pattern 47

48 Approaches (currently under progress) OpenCV :: AdaBoost Random Forest YOLO Use Darknet Library (found at: Design New CNN 48

49 Thanks To MRL Humanoid (MRL-HSL) MRL Biped Lab. (MRL-SPL) MRL3D Soccer Simulation Team 49

50 Any Questions? The code can be found on my github at: The documentation is available on my profile at: You can also reach my by via: 50