Self-driving cars need to perceive the world as humans do, when they drive. Humans use their eyes to figure out how fast they go, where the lane lines are, and where are the turns. Car does not have eyes but self-driving cars can use cameras and other sensors to keep the similar function. So what does cameras are seeing as we drive down the road is not the same as what we perceive. We have to teach the car how to perceive lane lines and turns. Hence, naturally, one of the first things we would like to do in developing a self-driving car is to automatically detect lane lines using an algorithm.
Here we are using OpenCV_A comprehensive Computer Vision library to understand captured images from the car and translate them to mathematical entities so we can show the car where the lane line is to follow. The algorithm I have chosen has multiple steps for manipulating each frame image in order to reduce or better said eliminate noises as much as possible. Since the car does not need to see the trees or clouds in the sky for detecting lane lines. In addition to reducing noises we try to highlight lane lines as much as possible by highlighting them using some image processing algorithms.
But the question is what features of lane lines do we need to highlight? Well, we can leverage following features to best identify various lane line in the image:
- position of the image
The line detection algorithm has multiple steps and the most important thing is to tweak all the required parameters well enough to not to loose any valuable pixel in the image.
The first step in detecting lane lines is understanding how images are represented. A 2D image can be represented as a rectangular grid, composed of many square cells, called pixels. Just like the black and white squares on the chessboard, pixels are nicely aligned in straight lines, both horizontally and vertically. We will refer to the horizontal ones as rows and to the vertical ones as columns. It is easy to see that a chessboard has 8 rows and 8 columns. But an image can have many rows and columns and we can find that out by using OpenCV image.shape . This returns a tuple of number of rows, columns and channels (if image is color).
>> image = cv2.imread('kitten.jpg') >> print(image.shape) (342, 548, 3)
If you want to know how many pixel this image has you can use img.size which returns the number of pixel in the image.
Each pixel in the image is represented in a 3D space as (R, G, B) values. Each of these values is in range [0-255]. So with this explanation we saw that images are basically tensors with different number of rows, columns and elements per each color channel.
Understanding of mathematical features of an image and its representation is a great help solving computer vision challenges and understanding image processing algorithms.
Lane line detection algorithm
Now let's describe the used algorithm to detect lane lines step by step.
Assume we have the above image captured by our car in a highway.
As you see lanes are either white or yellow on the streets. So we need to identify both.First we need to convert our image shape from a tensor (A, B, C) to a Matrix (A, B) to be able to only deal with raw pixels. In this case yellow and white considered both the same. In order to achieve that we can use OpenCV GrayScale method.
def grayscale(img): return cv2.cvtColor(img, cv2.COLOR_BGR2GRAY)
- . pixel Filtering As an enhancement here we also can identify only pixels with white or yellow color which helps out to filter out more pixels.
def select_white_yellow(image): converted = cv2.cvtColor(image, cv2.COLOR_RGB2HLS) # white color mask lower = np.uint8([ 0, 200, 0]) upper = np.uint8([255, 255, 255]) white_mask = cv2.inRange(converted, lower, upper) # yellow color mask lower = np.uint8([ 10, 0, 100]) upper = np.uint8([ 40, 255, 255]) yellow_mask = cv2.inRange(converted, lower, upper) # combine the mask mask = cv2.bitwise_or(white_mask, yellow_mask) return cv2.bitwise_and(image, image, mask = mask)
- . Gaussian Blur Now we need to apply a smoothing function to reduce image noise and detail. This also helps to smooth out edges in the image before applying Canny edge detection algorithm which is the next step.
def gaussian_noise(img, kernel_size): return cv2.GaussianBlur(img, (kernel_size, kernel_size), 0)
- . Canny Edge Detection This approach using differential values of the image to detect boundaries which are between three regions. Those regions are identifies by high and low thresholds which we feed to the function.
def canny(img, low_threshold, high_threshold): return cv2.Canny(img, low_threshold, high_threshold)
Here to maximize the edges I did dialated the image too.
- . Masked Image Now we have to narrow down our analysis to a section of image that lane lines are. In order to to do apply a trapezoidal mask over its edges. Here as an enhancement I applied two different masks one for right and one for left lane to have better detection specially on turns and curvy lanes.
def region_of_interest(img, vertices): #defining a blank mask to start with mask = np.zeros_like(img) #defining a 3 channel or 1 channel color to fill the mask with depending on the input image if len(img.shape) > 2: channel_count = img.shape # i.e. 3 or 4 depending on your image ignore_mask_color = (255,) * channel_count else: ignore_mask_color = 255 #filling pixels inside the polygon defined by "vertices" with the fill color cv2.fillPoly(mask, vertices, ignore_mask_color) #returning the image only where mask pixels are nonzero masked_image = cv2.bitwise_and(img, mask) return masked_image
. Hough Transform
So far we did a great job identifying lane lines in the picture, now we need to draw them to show the machine where the lanes are. In order to draw lines in images or video frames we use a technique called Hough Transform which is basically determines for each given pixel in the picture if there is a straight line passing through the pixel. The algorithm uses 5 parameters which are rho, theta, min_votes, min_line_length and max_line_gap. You can read more about Hough transform algorithm here.
lines = cv2.HoughLinesP(img, rho, theta, threshold, np.array(), minLineLength=min_line_len,maxLineGap=max_line_gap) #line_img = np.zeros(img.shape, dtype=np.uint8) #draw_lines(line_img, lines) #return line_img return lines
We draw the lines separately for left and right side. I did tweak this function a little bit for applying some more filtering on lines and adjusting line's slopes. I am going go cover those in a the sext section.
- . Adjustments to improve lane detection
a. Filtering slopes
Here I have applied some enhancements to draw lines more efficiently. When I first applied hough transform I got so many other lines which their slopes were different than the lane line and I have to filter those out. Following function looks through all the hough transformed lines and filter those out based on the defined min and max slopes. This will help specifically with curved lines which slope changes are significant.
def slope_filter(lines_array, positive, min_slope, max_slope): slopes = np.apply_along_axis(lambda row: (row - row) / (row - row), 2, lines_array) if positive: slopes[slopes > max_slope] = 0 slopes[slopes < min_slope] = 0 lines_array = np.array(lines_array[np.where(slopes > 0)]) else: slopes[slopes < -max_slope] = 0 slopes[slopes > -min_slope] = 0 lines_array = np.array(lines_array[np.where(slopes < 0)]) return lines_array
b. Linear Regression
Linear Regression comes handy when identifying curved lines. It uses standard error of the estimate is a measure of the accuracy of predictions. Recall that the regression line is the line that minimizes the sum of squared deviations of prediction (also called the sum of squares error). The standard error of the estimate is closely related to this quantity and is defined below:
where σest is the standard error of the estimate, Y is an actual score, Y' is a predicted score, and N is the number of pairs of scores. The numerator is the sum of squared differences between the actual scores and the predicted scores.
In summary if we have a line, y = mx + c, through some noisy data-points, By examining the coefficients, we see that the line should have a gradient of roughly g1 and cut the y-axis at, more or less, y0. We can rewrite the line equation as y = Ap, where A = [[x 1]] and p = [[m], [c]]. Now use lstsq to solve for p.
In a simpler language linear regression can draw a best line regarding a given number of points so that the Root Mean Squared Error is minimized.
As you must realized linear regression can do the same identifying the best line among all the hough lines and efficenizing the slope by reducing RMSE.
def lines_linreg(lines_array): ### Select the 0th and 2nd index which will provide the xval and reshape to extract x values x = np.reshape(lines_array[:, [0, 2]], (1, len(lines_array) * 2)) ### Select the 1st and 3rd index which will provide the yval and reshape to extract 7 values y = np.reshape(lines_array[:, [1, 3]], (1, len(lines_array) * 2)) A = np.vstack([x, np.ones(len(x))]).T m, c = np.linalg.lstsq(A, y) x = np.array(x) y = np.array(x * m + c) return x, y, m, c
- . Weighted Image In this step we only need to overlay hough image and original image to display lines we call the resulted image weighted image.
def weighted_img(img, initial_img, α=0.8, β=1., λ=0.): return cv2.addWeighted(initial_img, α, img, β, λ) color_with_lines = np.zeros(image.shape, dtype=np.uint8) color_with_lines = draw_lines(color_with_lines, lines_left, lines_right, [255, 0, 0], 10) result = weighted_img(color_with_lines, image)
One potential shortcoming would be what would happen when lane detection is happening on roads with low visibility like driving at night or in rainy days. What if the road does not have any line? like when you drive in desert.
Also we need more images and videos to try and enhance the algorithm with.
One thing I really was not satisfied with here was not being able for the car to adapt to data and take decisions. This algorithm is like telling the car what to see all the time.
- use Deep learning to make the car adapt with road.
- drawing the whole lane instead of only marking lane lines