This project was coded in 2015 for self-driving project using stereo vision. I had conducted self-driving based on stereo vision. This code doesn't include detection & tracking. In this code, My point was preprocessing for exact obstacle detection.
- I used Block Matching(BM) Algorithm for a disparity map.
- I separated ground and unnecessary parts(above the height of a vehicle) by using v-disparity method.
- I coded occupancy grid which is one of the General Object Detection(GOD) methods.
- I tried stixels method.
-
Dataset
Daimler open dataset (TestData_c0_part1.tgz, TestData_c1_part1.tgz)
Windows 8.1(x64), Visual studio 2013, OpenCV 2.4.9
CPU : i3-4005U 1.72GHz, Memory : 6GB
I used Block Matching(BM) Algorithm in this project. but you can use (Semi-Global Matching)SGM Algorithm by changing alg = 0 to 1. As you can see, the SGM prints out better disparity map but I used a BM because of real-time processing. If your machine is good enough, I recommend SGM.
I separated ground using V-disparity. And I used RANSAC algorithm for line fitting. I had tried separating ground with geometric method previously, but v-disparity result is better.
For more details about V-disparity, check this paper (Real Time Obstacle Detection in Stereovision on Non Flat Road Geometry Through ”V-disparity”Representation).
I also erased unnecessary parts(above the height of a vehicle) in the same way as V-disparity.
This is a occupancy grid whis is one of the General Object Detection(GOD) methods.
Red color means ground, Yellow means obstacles within 13m, Blue means obstacles over 13m
I also tried Stixels methods simply. This method is proposed from Daimler group. For more details, check this paper(The Stixel World - A Compact Medium Level Representation of the 3D-World).
video result (Click for full video)
