latent-pixel / lanedetection

Lane detection and turn prediction using probabilistic Hough transforms.

Lane Detection and Turn Prediction

Lane detection algorithms are widely employed on vehicles to identify and track lanes on a road from images/video feed that is captured by mounted cameras. Lane detection is an important component of the Advanced Driver Assistance Systems (ADAS) suite in modern vehicles.

The goal of this project is to develop a simple lane detection algorithm that could be used to first detect lanes from a video feed and then track the detected lanes to predict the vehicle turn.

Dependencies

This package was built using Python 3.7 on Ubuntu 20.04. OpenCV can be installed by following the instructions on . Other packages used include scipy and matplotlib, which are pretty easy to install using the pip package manager:

pip install <package_name>

Then, download the package using

git clone [email protected]:latent-pixel/LaneDetection.git

1. Histogram Equalization

To get a taste of the image preprocessing we might need when dealing with real-world data, we will first look at Histogram Equalization methods that are useful in "balancing" out image intensities. The goal is to make the image more visually appealing or suitable for further image processing tasks.

Running the package

To run the package, use the following command:

python code/HistEqualization.py -i <input_file_path> -t <norm or adap> -s <save_path>

• Histogram type is either "norm" or "adap" for traditional histogram equalization and adaptive histogram equalization, respectively.

Sample Output

Upon running the package, we get a .avi file which contains the histogram-equalized input.

Original	Traditional Histogram Equalization	Adaptive Histogram Equalization

2. Lane Detection

There are several approaches and algorithms for lane detection, each with its own strengths and limitations. For this project, we make use of the Probabilistic Hough Transform to tackle the problem.

Running the package

To run the package, use the following command:

python code/LaneDetection.py -i <input_file_path> -s <save_path>

• Note that this is a relatively simple (classical) lane detection solution, you might have to play around with thresholding and Hough Transform parameters to get the desired output for another problem.

Sample Output

Upon running the package, we get a .avi file which contains the frames with lane detection.

3. Turn Prediction

The goal is to detect curved lanes and predict the turn depending on the curvature: left or right. For this, we need compute the radius of curvature of the road by fitting a curve through the detected lane lines.

Running the package

To run the package, use the following command:

python code/LaneDetection.py -i <input_file_path> -s <save_path>

Sample Output

Upon running the package, we get a .avi file which contains the frames with lane detection.

Figure 1 (Clockwise from top-left): Original frame, thresholded frame, Bird's eye view with curve fitting, Turn prediction with lane overlay.

Sometimes, we face bad input frames with color and exposure issues, one such frame is shown below. To solve such problems, we use weighted average approach where the curve-fitting parameters that were historically stored weigh in heavily and predict the lane lines.

Figure 2 (Clockwise from top-left): Original frame, thresholded frame, Bird's eye view with curve fitting - notice the problems with contrast, Turn prediction with lane overlay.