1. Segmentation of Floor in Corridor Images for Mobile Robot Navigation Yinxiao Li Clemson University Committee Members: Dr. Stanley Birchfield (Chair) Dr. Adam Hoover Dr. John Gowdy

2. Motivation Goal: Segment the floor in a single corridor image Why? obstacle avoidance mapping autonomous exploration and navigation

3. Motivation Why? obstacle avoidance mapping autonomous exploration and navigation Goal: Segment the floor in a single corridor image

4. Outline Previous Work Detecting Line Segments Score Model for Evaluating Line Segments –Structure Score –Bottom Score –Homogeneous Score Experimental Results Extension

5. Outline Previous Work Detecting Line Segments Score Model for Evaluating Line Segments –Structure Score –Bottom Score –Homogeneous Score Experimental Results Extension

6. Previous Work Free space detection Combination of color and histogram (Lorigo 1997) Optical flow (Stoffler 2000, Santos-Victor 1995) Stereo matching (Sabe 2004) Floor detection Apply planar homographies to optical flow vectors (Kim 2009, Zhou 2006) Stereo homographies (Fazl-Ersi 2009) Geometric reasoning (Lee 2009) Limitations Multiple images for motion Multiple cameras for stereo Require different colors for floor and wall Computational efficiency Calibrated cameras Assume ceiling visible Our contribution: Segment the floor in real time using a single image captured by a low-height mobile robot

7. Challenging problem: Reflections Also: variety of poses (vanishing point, ceiling not always visible) sometimes wall and floor color are nearly the same

8. Douglas-Peucker Algorithm Purpose: Reduce the number of points in a curve (polyline) Algorithm: 1. Connect farthest endpoints 2. Repeat 1.Find maximum distance between the original curve and the simplified curve 2.Split curve at this point Until max distance is less than threshold David Douglas & Thomas Peucker, "Algorithms for the reduction of the number of points required to represent a digitized line or its caricature", The Canadian Cartographer 10(2), 112–122 (1973)

9. Douglas-Peucker Algorithm Purpose: Reduce the number of points in a curve (polyline) Algorithm: 1. Connect farthest endpoints 2. Repeat 1.Find maximum distance between the original curve and the simplified curve 2.Split curve at this point Until max distance is less than threshold

10. Douglas-Peucker Algorithm Purpose: Reduce the number of points in a curve (polyline) Algorithm: 1. Connect farthest endpoints 2. Repeat 1.Find maximum distance between the original curve and the simplified curve 2.Split curve at this point Until max distance is less than threshold

11. Douglas-Peucker Algorithm Purpose: Reduce the number of points in a curve (polyline) Algorithm: 1. Connect farthest endpoints 2. Repeat 1.Find maximum distance between the original curve and the simplified curve 2.Split curve at this point Until max distance is less than threshold

12. Douglas-Peucker Algorithm Purpose: Reduce the number of points in a curve (polyline) Algorithm: 1. Connect farthest endpoints 2. Repeat 1.Find maximum distance between the original curve and the simplified curve 2.Split curve at this point Until max distance is less than threshold

13. Douglas-Peucker Algorithm Purpose: Reduce the number of points in a curve (polyline) Algorithm: 1. Connect farthest endpoints 2. Repeat 1.Find maximum distance between the original curve and the simplified curve 2.Split curve at this point Until max distance is less than threshold

14. Detecting Line Segments (LS)

15. 1.Compute Canny edges

16. Detecting Line Segments (LS) 1.Compute Canny edges 2.Douglas-Peucker algorithm to detect line segments Vertical LS: slope within ±5° of vertical direction Horizontal LS: Slope within ±45° of horizontal direction

17. Detecting Line Segments (LS) 1.Compute Canny edges 2.Douglas-Peucker algorithm to detect line segments Vertical LS: slope within ±5° of vertical direction Horizontal LS: Slope within ±45° of horizontal direction 3. Pruning line segments (320x240) Vertical LS: minimum length 60 pixels Horizontal LS: minimum length 15 pixels Vanishing point (height selection): The vanishing point is computed as the mean of the intersection of pairs of non- vertical lines. It is used for throwing away horizontal line segments coming from windows, ceiling lights, etc.

18. Outline Previous Work Detecting Line Segments Score Model for Evaluating Line Segments –Structure Score –Bottom Score –Homogeneous Score Experimental Results Extension

19. Score Model is assigned to each horizontal line Structure Score Bottom Score Homogeneous Score Weights horizontal line

20. Score Model – Structure Given a typical corridor image 1.|  I(x,y)| > T 1 (gradient magnitude) (How to set threshold T 1 ?) |  I(x,y)| > T 1

21. Score Model – Structure Given a typical corridor image 1.|  I(x,y)| > T 1 (gradient magnitude) (How to set threshold T 1 ?) 2.I(x,y) < T 2 (image intensity) (How to set threshold T 2 ?) |  I(x,y)| > T 1 and I(x,y) < T 2

22. Score Model – Structure Given a typical corridor image 1.|  I(x,y)| > T 1 (gradient magnitude) (How to set threshold T 1 ?) 2.I(x,y) < T 2 (image intensity) (How to set threshold T 2 ?) We applied SVM to 800 points in 200 images: |  I(x,y)| > T 1 and I(x,y) < T 2

23. Score Model – Structure Given a typical corridor image 1.|  I(x,y)| > T 1 (gradient magnitude) (How to set threshold T 1 ?) 2.I(x,y) < T 2 (image intensity) (How to set threshold T 2 ?) Approximate using two thresholds: |  I(x,y)| > T 1 and I(x,y) < T 2 T2T2 T1T1

24. Score Model – Structure Given a typical corridor image 1.|  I(x,y)| > T 1 (gradient magnitude) 2.I(x,y) < T 2 (image intensity) 3.Now threshold original image using T LC (average gray level of pixels satisfying #1 and #2) I(x,y) < T LC

25. Score Model – Structure Given a typical corridor image 1.|  I(x,y)| > T 1 (gradient magnitude) 2.I(x,y) < T 2 (image intensity) 3.Now threshold original image using T LC (average gray level of pixels satisfying #1 and #2) 4.Compute the chamfer distance between the line segments and structure blocks

26. Score Model – Structure Ridler-Calvard Otsu Our method Our method is able to remove the spurious pixels on the floor caused by reflections or shadows MethodR-COtsuOurs Correctness62%66%82% Comparison of different threshold methods

27. Score Model – Bottom 1.Connect the bottom points of consecutive vertical line segments to create “bottom” wall-floor boundary

28. Score Model – Bottom 1.Connect the bottom points of consecutive vertical line segments to create “bottom” wall-floor boundary

29. Score Model – Bottom 1.Connect the bottom points of consecutive vertical line segments to create “bottom” wall-floor boundary 2.Compute the distance of each horizontal line segment to the “bottom” wall-floor boundary (red circled horizontal line segments are more likely on the real wall-floor boundary)

30. Score Model – Homogeneous Idea: The floor tends to have larger regions (due to decorations, posters, windows on the wall). Algorithm: 1.Color-based segmentation of the image (using Felzenszwalb- Huttenlocher’s minimum spanning tree algorithm) 2.For each horizontal line segment, compute size of region just below segment size of largest segment

31. Score Model – Homogeneous Idea: The floor tends to have larger regions (due to decorations, posters, windows on the wall). Algorithm: 1.Color-based segmentation of the image (using Felzenszwalb- Huttenlocher’s minimum spanning tree algorithm) 2.For each horizontal line segment, compute size of region just below segment size of largest segment

32. Segmenting the floor Normalize the scores and sum Threshold the final score: Connect the remaining line segments and extend the endpoints to the edges of the image

33. Evaluation Criterion Wall- floor Boundary –Green line: Detected wall-floor boundary –Red line: Ground truth

34. Evaluation Criterion Wall- floor Boundary –Green line: Detected wall-floor boundary –Red line: Ground truth Area –Blue shade area: misclassified area –Red shade area: ground truth floor area Error Rate Segmentation is considered successful if r err < 10%

35. Outline Previous Work Detecting Line Segments Score Model for Evaluating Line Segments –Structure Score –Bottom Score –Homogeneous Score Experimental Results Extension

36. Sample Results 89.1% success on database of 426 images:

37. Sample Results Images downloaded from the internet:

38. Sample Results Some successful results on failure examples from Lee et al. 2009 Original Image Lee’s Ours D. C. Lee, M. Hebert, and T. Kanade. “Geometric Reasoning for Single Image Structure Recovery”. IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2009.

39. Sample Results Failure examples Checkered floor Textured wall Dark image Bright lights

40. Sample Videos Video clip 1

41. Sample Videos Video clip 2

42. Outline Previous Work Detecting Line Segments Score Model for Evaluating Line Segments –Structure Score –Bottom Score –Homogeneous Score Experimental Results Extension

43. Floor Segmentation Algorithm on Low-Res Images Apply floor-segmentation algorithm to varying image resolutions – Simply change one parameter: Minimum length of a horizontal line segment: Combine with Corridor Orientation line estimation* * V. N. Murali and S. T. Birchfield, “Autonomous Exploration Using Rapid Perception of Low-Resolution Image Information”, Autonomous Robots (In review)

44. Results for varying resolutions Wall-floor Boundary

45. Results for minimalistic geometry Minimalistic geometry estimation* *V. Murali, Y. Li, and S.T. Birchfield, “Extracting Minimalistic Corridor Geometry from Low-Resolution Images”, IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), 2011.(In review)

46. Sample Videos 1

47. Sample Videos 2

48. Conclusion Image-Based Floor Segmentation - Edge-based approach to segmenting floors in corridors - Correctly handles specular reflections on floor - Nearly 90% of the corridor images in our database can be correctly detected. - Speed: approximately 7 frames / sec - Extension - Three line “geometry”---enough for basic robot navigation

49. Future Work Image-Based Floor Segmentation - Speed up the current high-res algorithm - Improving score model by add more visual cues (highly textured floors, low resolution, or dark environment) - Adapt weights by Adaboost training - Use floor segmentation for mapping

50. Thanks! Questions? Segmentation of Floor in Corridor Images for Mobile Robot Navigation ---By Yinxiao Li