After receiving the test image (or a frame of image from a video clip) a color filter that filters out everything expect for white and yellow is applied. Specifically, the filter uses a RBG color thresholds for white color and HLS color thresholds for color yellow:
# color thresholds for white and yellow
white_lo = [100, 100, 200]
white_hi = [255, 255, 255]
yellow_lo = [20, 120, 80]
yellow_hi = [45, 200, 255]A HLS color filter, rather than a RGB one, is employed for color yellow because GRB filter also allows white spectrum and does not give robust output. For example, this is a sample result after the color filter is applied:
Now it is possible to convert the image into grayscale and apply a Gaussian filter to blur the image, a preparation step for Canny Edges Detection. After Canny Edges Detection the image looks like this:
Our region of interest is only the trapezoid area to the center bottom of the image, so after masking the image with this region of interest, only the contour of two lines are visible:
After Hough Transform, the canny edges are interpreted as voted discrete x-y coordinates and it is crucial to turn these scattered points into two lines. We adopt linear regression technique and classify the points into two groups, left_x, left_y and right_x, right_y, determined by their x-coordinate. This enables linear regression line fit on the two categories:
left_slope, left_int = np.polyfit(left_x, left_y, 1)
right_slope, right_int = np.polyfit(right_x, right_y, 1)Furthermore, a quasi-memorization mechanism makes smooth rendering possible. Four global variables, SAVED_LB, SAVED_LT, SAVED_RB, SAVED_RT, memories and updates the weighted x-coordinates of 4 endpoints so far with a weight of 0.5. After all processes are done, the output image has two marked lane lines on the original input image:
- The pipeline has problem identifying curves, since HSL and RGB filters does not return a robust output with curved lines, and the linear regression technique does not approximate curved lines.
- Not possible to dynamically adjust region of interest