p1-lanelines on Github


OpenCV Toolbox

OpenCV is an image processing toolbox originally developed in C++. In Python, an OpenCV image is a numpy array (2D or 3D depending on the type of image). The figure below depicts the coordination used in OpenCV. For example, if we have a numpy array img describing an OpenCV image, then img[0,0] stores the data for top left pixel having the coordinate (x=0,y=0). Following this system, the bottom right corner img[-1,-1] has the coordinate (x=img.shape[1]-1, y=img.shape[0]-1) in OpenCV.

OpenCV Image

OpenCV provides various tools for us to “get our hand dirty” with images. Generally, there are two main groups: image drawing and image transformation. Drawing on images in OpenCV is quite simple and straight forward (except for the ellipse :smile:) as most drawing function is in the form:

import cv2
cv2.(img_to_draw_on, starting_point, **others_arguments)

Line Detection Pipeline

The main tool to detect lines in an image is a technique named Hough Transformation. It is called transformation because it transforms the representation of a line to a pair of line angle and line distance to the origin. The detail and tutorial of Hough Transformation is provided here. To reduce the computation and increase the accuracy of the line detection pipeline, we focus only on the area in front of a car. The following list describes our pipeline:

  1. Extract the region of interest (ROI). In this case, the ROI is a trapezoid in front of the car.
  2. Create a color mask. Lucky for us, the lane lines of interest only have white or yellow colors. Therefore, extracting only yellow and white color will greatly reduce the computation for unwanted objects in the image.
  3. Canny edge detection. Performing the Canny edge detection algorithm on the color filtered image greatly reduces number of points needed to process in the next step.
  4. (Probabilistic) Hough line transform. Given a set of points from edge detection, we detect lines using Hough transformation.
  5. Split lines into left and right set. The output of Hough line transformation is a set of lines, represented by two points (x1,y1) and (x2,y2). For each pair of points in the returned set, we split them into left lane points and right lane points by its coefficient.
  6. Fitting lines to left and right points. We get two lines for left lane and right lane by fitting a line to each set of points. The output of OpenCV’s line fitting algorithm is a 4-tuple: (x0, y0, vx, vy), where (x0,y0) is a point on the line and (vx,xy) is the line’s co-linear vector.
  7. Drawing lines onto the image. To draw a line onto an image, we need two points (start and end). We can compute these points for drawing from the output of step 6 and a specified drawing zone (the trapezoid ROI for example).

Reflection

Color space and region of interest

The first pipeline we came up with in this project simply converts the input to gray scale, detect edges by Canny algorithm, and then draw all line that Hough transformation returns. This approach is unstable since it depends heavily on the high and low threshold of the Canny algorithm. Although using a slight Gaussian blur on the gray scale image can reduce noise and improve the quality of the pipeline, imperfection on the road can potentially disturb the pipeline’s robustness.

The unused information in our first pipeline is: 1. The region of interest, and 2. The color of the lane lines. By extracting the region of interest, we eliminate unnecessary computation:

Focus region

In addition to extracting the region of interest, we also filtered out unwanted colors. By default, the image output of cv2.imread is a GBR image. This color representation makes it hard to filter a certain color since all three values (G,B,R) of a pixel represents color. Therefore, we convert the image to HSV color space (Hue, Saturation, Value). In HSV images, a pixel contains the color (hue), the “amount” of that color (saturation), and its brightness (value). This color representation enable us to specify the colors we want to extract. To exact colors, the rule of thumb is to range ±10 in the hue value as following:

hue_range = 10  # Increase for wider color selection
# rgb_color is the (R,G,B) tuple value of the color we want to filter
pixel = np.uint8([[rgb_color]])  # One pixel image
hsv_pixel = cv2.cvtColor(pixel, cv2.COLOR_RGB2HSV)  # Convert to HSV
hue = hsv_pixel[0,0,0]  # Get the hue value of the input (R,G,B)
lowb = np.array((hue-hue_range, 100, 100), dtype=np.uint8)
upb = np.array((hue+hue_range, 255, 255), dtype=np.uint8)
return lowb, upb  # Lower and upper bound for color filtering

To exact black or white color, the code is different since it depends on the saturation and value rather than the hue.

sensitivity = 30
lowwhite = np.array((0,0,255-sensitivity), dtype=np.uint8)
upwhite = np.array((255,sensitivity,255), dtype=np.uint8)
return lowwhite, upwhite  # Lower and upper bound for color filtering

After selecting only the region of interest and the colors, we have the following result:

Focus region

The image above is a binary image which can be used as a mask to extract the lane lines from the original image. The example of our lane lines detection on static image is shown below.

Result on image

Buffered pipeline

The pipeline showed in the previous session performs well on test images and videos. However, with the challenge video, it failed to detect the lane lines for some brief moments when the lighting varies. Furthermore, in all videos, the lane lines between frame doesn’t have smooth transitions. To address this problem, we have several approaches:

  1. Limit the movement of lines between frames. We specify a limit for the displacement of two lines between adjacent frames. The next frame’s line is computed as:
  2. Store previous lines in a fixed-size buffer, add new line to the buffer for every frame. The output is the weighted average of all the lines in the buffer.
  3. Similar to the second approach, but instead of storing line points, we store the lines’ co-linear vectors. The next line’s co-linear vector is the weighted average of the vectors stored in the buffer.

The videos result for each of the approach will be updated soon. TODO: Upload videos.

We have some minor bugs during the implementation of of the buffered pipeline. Firstly, when the buffer is empty, the pipeline should not draw the line. In one of our implementation, a default line is drawn when the buffer is empty, this design decision makes it hard to debug the program. Secondly, when no line is detected from the frame, the algorithm should still return a line from the buffer. However, if there are many “no line” frames, there is a chance that there isn’t any lane lines on the road. Our current buffer implementation hasn’t taken care of this situation.

Unnecessary operations

For the current testing data (images and videos), extracting the region of interest and lane line colors is enough for line detection.

Result on image

As the picture above has shown, only yellow (left) and white (right) color filter is enough to give us a substantially clear image of lane lines. This output here can be put directly to the Hough Line detection (without masking with the original image or Canny edge detection) to obtain the lane lines. At this stage, we don’t know if performing Canny edge detection is necessary (i.e. makes the pipeline more robust) :confused:.

Thanks for reading! :sunny: