Lane detection using regression.
At first bird-eye view is obtained and then histogram is calculated.
After calculating the histogram a moving average sliding window is used to
calculate the good points
The points are used for regression to find a polynomial.
Steps followed:
Gaussian Blur: In, a Gaussian blur image processing (also known as Gaussian smoothing) is the result of blurring an image by a (named after Gaussian function mathematician and scientist). The visual effect image noise of this blurring technique is a smooth blur resembling that of viewing the through a translucent screen. Gaussian Blur is used to make the edges smoother.
Gray scaling: The images should be converted into gray scaled ones in order to detect shapes (edges) in the images. This is because the canny edge detection measures the magnitude of pixel intensity changes or gradients (more on this later).
Canny edge detection: Canny edge detection which make use of the fact that edges has high gradients (how sharply image changes — for example, dark to white)
Region of interest: When finding lane lines, we are not interested in other environmental disturbances .So we mask the other regions as per our requirement.
Perspective transform: We now need to define a trapezoidal region in the 2D image that will go through a perspective transform to convert into a bird’s eye view. It’s the change in the perspective when we see the image from side and when we see it as top-down view.
The curve lines looks almost straight after the transform as can be seen in the below image.
Histogram: We then compute a histogram of our binary thresholded images in the y direction, on the bottom half of the image, to identify the x positions where the pixel intensities are highest.
Here the frame is divided into two parts, left and right sides and that is how we get the exact indices of the frame. The operation is done on both sides to detect the lane on both the sides.
As it can be seen in the below image the histogram is quite high at the position where lane is detected.
Sliding window: Since we now know the starting x position of pixels (from the bottom of the image) most likely to yield a lane line, we run a sliding windows search in an attempt to “capture” the pixel coordinates of our lane lines.
Next, we compute a second degree polynomial, via numpy’s polyfit, to find the coefficients of the curves that best fit the left and right lane lines.
The detected lane can be seen below:
