Stereo Vision

Introduction

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.

Summary

• 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.

Installation & Environment

Installation

• Dataset

  Daimler open dataset (TestData_c0_part1.tgz, TestData_c1_part1.tgz)

Environment

software

Windows 8.1(x64), Visual studio 2013, OpenCV 2.4.9

hardware

CPU : i3-4005U 1.72GHz, Memory : 6GB

Stereo matching Algorithm

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.

Separate ground (searching free space)

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).

Except unnecessary parts (above the height of a vehicle)

I also erased unnecessary parts(above the height of a vehicle) in the same way as V-disparity.

Occupancy grid (top view)

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

Stixels

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).

Result

video result (Click for full video)