Monocular Camera-based Real-time Image Recognition for Autonomous Vehicle

Abstract

Homo Faber, refers to humans as controlling the environments through tools. From the beginning of the world, humans create tools for chasing the convenient life. The desire for the rapid movement let the human ride on horseback, make the wagon and finally make the vehicle. The vehicle made humans possible to travel the long distance very quickly as well as conveniently. However, since human being itself is imperfect, plenty of people have died due to the car accident, and people are dying at this moment. The research for autonomous vehicle has been conducted to satisfy the humans desire of the safety as the best alternative. And, the dream of autonomous vehicle will be come true in the near future. For the implementation of autonomous vehicle, many kinds of techniques are required, among which, the recognition of the environment around the vehicle is one of the most fundamental and important problems. For the recognition of surrounding objects many kinds of sensors can be utilized, however, the monocular camera can collect the largest information among sensors as well as can be utilized for the variety of purposes, and can be adopted for the various vehicle types due to the good price competitiveness. I expect that the research using the monocular camera for autonomous vehicle is very practical and useful. In this dissertation, I cover four important recognition problems for autonomous driving by using monocular camera in vehicular environment. Firstly, to drive the way autonomously the vehicle has to recognize lanes and keep its lane. However, the detection of lane markings under the various illuminant variation is very difficult in the image processing area. Nevertheless, it must be solved for the autonomous driving. The first research topic is the robust lane marking extraction under the illumination variations for multilane detection. I proposed the new lane marking extraction filter that can detect the imperfect lane markings as well as the new false positive cancelling algorithm that can eliminate noise markings. This approach can extract lane markings successfully even under the bad illumination conditions. Secondly, the problem to tackle, is if there is no lane marking on the road, then how the autonomous vehicle can recognize the road to run? In addition, what is the current lane position of the road? The latter is the important question since we can make a decision for lane change or keeping depending on the current position of lane. The second research is for handling those two problems, and I proposed the approach for the fusing the road detection and the lane position estimation. Thirdly, to drive more safely, keeping the safety distance is very important. Additionally, many equipments for the driving safety require the distance information. Measuring accurate inter-vehicle distance by using monocular camera and line laser is the third research topic. To measure the inter-vehicle distance, I illuminate the line laser on the front side of vehicle, and measure the length of the laser line and lane width in the image. Based on the imaging geometry, the distance calculation problem can be solved with accuracy. There are still many of important problems are remaining to be solved, and I proposed some approaches by using the monocular camera to handle the important problems. I expect very active researches will be continuously conducted and, based on the researches, the era of autonomous vehicle will come in the near future.