1 of 15

Line Following on the Racecar Using Computer Vision

RSS Team 15

An Bo Chen, Rachel Lu, Lauren Carethers, Brian Li, and Claire Lu

2 of 15

Achieving Line-Following Using Computer Vision

2

Presenter(s): Rachel

5x

RSS Team 15

The main goal of the labs is to achieve an efficient and effective autonomous racecar that can handle various situations. In the previous lab, we build an autonomous wall follower. However, this wall follower program will not work effectively in certain situations. Therefore. this lab focuses on building upon the autonomous wall follower from the previous lab by introducing an autonomous line follower. To achieve this line follower program, we first used color segmentation to detect the cones and lines from computer vision, homography transformation to convert pixel coordinates to world coordinates, and a parking controller to enable the racecar to stop within a certain distance of a specified color. Before finalizing on using color segmentation for our computer vision, we experimented with using Scale Invariant Feature Transform or SIFT which will be further explained by Lauren.

3 of 15

SIFT is Effective on the CITGO Dataset

Score statistics

Mean: 0.51
Median: 0.71
Standard deviation: 0.37

3

Presenter(s): Lauren

RSS Team 15

For object detection we tested and compared a SIFT algorithm and Template Matching algorithm.

Focusing first on SIFT, we learned that is was essentially a feature extraction method that reduces an image’s content to a set of points that are used to detect similar patterns in other images.

And upon implementing and testing our SIFT algorithm, we found that it generally performed well for the CITGO sign test cases with a score median of 0.71 and score standard deviation of 0.37, but it performed poorly for the cone and map test cases, only finding between 0 to 7 detected matches between the template and source images. Some pros we realized with SIFT is that the algorithm performances well independent of the scale, rotation, translation, illumination, and blur of the template and source image, has locality, so features are not affected by noise and clutter, is distinct such that features can be compare to large data sets of objects, it can generate many features even from small objects and it also runs relatively efficiently and can be compared to real-time performance.

But its biggest cons are that it can be computationally heavy and mathematically complicated

_____________________________

Scale Invariant Feature Transform (SIFT)

4 of 15

Template Matching is Effective on the Stata Dataset

Score statistics

Mean: 0.10
Median: 0.0
Standard deviation: 0.16

4

Presenter(s): Lauren

RSS Team 15

For template matching, we found that the algorithm essentially works by sliding the template image across the source image to find matches, which means that changes in the size of the source or the template image will affect the ability of the algorithm to actually detect matches.

For this reason, we found that the Template Matching algorithm performed pretty well in the stata map test case as the template images in this test case could directly be compared to the source image as they were already in the same orientation as they would be found in in the source image.

But in the real-world, template images and source images will likely not have similar orientations and sizes all the time, and a template image may not exactly be contained fully in the source image, so template matching becomes a far less reliable and accurate object detection method compared to SIFT.

5 of 15

Color Segmentation is Effective for Identifying Cones

IOU Scores Statistics

Median: 0.79
Standard Deviation: 0.14

5

Presenter(s): An Bo

RSS Team 15

6 of 15

Homography Transformation of Coordinates to Pixels

6

Presenter(s): Rachel

Conversion between Pixel-frame and Robot-frame Coordinates

RSS Team 15

7 of 15

Implementing Pure Pursuit for the Parking Controller

7

Presenter(s): Brian

L = 1 for length of racecar

L_d = lookahead distance

α = angle difference between car’s heading and target point

RSS Team 15

8 of 15

Racecar Successfully Parks in front of Cone

8

Presenter(s): Claire

RSS Team 15

9 of 15

Calculated Error Converges to Zero

9

Presenter(s): Claire

RSS Team 15

10 of 15

Modifying the Parking Controller to Follow a Line

Decrease Parking Distance

Restrict field of view

Orange detection towards center of image

10

Presenter(s): An Bo

RSS Team 15

11 of 15

Racecar Follows an Orange Line

[videos / error graphs]

11

Presenter(s): Brian

5x

RSS Team 15

12 of 15

Parking Error Remains Within a Small Range

12

Presenter(s): Claire

RSS Team 15

13 of 15

Lessons Learned and Next Steps

Technical

Debugging with robot camera view
Improve homography calibration
Adapt color ranges to follow Johnson track

Team

Division of project
Communication among modules

13

Presenter(s): Rachel

RSS Team 15

14 of 15

Appendix:

14

RSS Team 15

15 of 15

Appendix: Template Matching for Map, different coefficients

TM_CCOEFF_NORMED

Median: 0.0
Standard deviation: 0.16

TM_CCOEFF

Median: 0.0
Standard deviation: 0.16

TM_SQDIFF, TM_CORR, TM_SQDIFF_NORMED, TM_CORR_NORMED,

Median: 0.0
Standard deviation: 0.0

15

RSS Team 15