392 0

A Vehicle Tracking Algorithm Using Laser Scanners Based on L-Shape Model Switching to Mitigate Vehicle Appearance Change

Title
A Vehicle Tracking Algorithm Using Laser Scanners Based on L-Shape Model Switching to Mitigate Vehicle Appearance Change
Author
김동철
Advisor(s)
선우명호
Issue Date
2017-02
Publisher
한양대학교
Degree
Doctor
Abstract
Recent attention on the development of highly safe and comfortable vehicles has been drawn not only by customers on the market but also governmental regulations. Those social concerns are a driving force to develop an intelligent vehicle which provides information regarding driving environment to the drivers and thus enables to assist safe and comfortable driving. In the near future, the intelligent vehicle technology will be realized and evolve into an autonomous car, that drives itself to a destination without human intervention, which is the ultimate goal of the intelligent vehicle. The development of autonomous car requires the state-of-the-art technologies in localization, perception, planning, control, and system management. Among those core technologies, the perception is an essential part which perceives road environment using various sensors to determine the behavior and motion of an autonomous car for the safe navigation. Among various sensors for the perception systems of autonomous car, a laser scanner is the most popular sensor because of its accurate and precise distance information which is also known as ‘point cloud’. However, this point cloud from the laser scanner only represents surface information facing the sensor and changes its appearance over time. Since this appearance change produces unexpected motion, the tracking system estimates position and velocity with large error. Many previous studies tried to solve the appearance change problem but their methods have some limitations such as complex structure, unstable tracking performance, and requirement of huge computational power. In this study, in order to overcome the appearance change problem, a vehicle tracking algorithm based on L-shape model switching is proposed. The proposed algorithm tracks L-shape feature which consists of corner point position, line length, and heading angle. The tracking system can successfully reduce the motion estimation error by tracking the corner point, because the corner point has invariant relative position to target vehicle appearance. In addition, following features of the proposed method can eliminate the limitations of the previous researches; Kalman filter based point tracking structure of the algorithm and the covariance adaptation of the L-shape model switching scheme. The first one provides simplicity of the algorithm and requires less computational power, the other makes tracking performance more reliable. The proposed algorithm is validated in simulation and real traffic experiments. In the simulation, basic functions and performances are analyzed with the laser scanner model. In the real road environment, the performance of the proposed method is evaluated. The evaluation of the performance is carried out by comparing the proposed method with a reference tracking system. As the reference system, the tracking system of IbeoLux is adopted, because the tracking performance of this system is the best among the commercially available laser scanners. In order to evaluate the performance, the proposed algorithm is tested in various scenarios. There are two kinds of test scenarios to validate functionality of the algorithm and evaluate performance of the algorithm. The functional validation scenarios are same as the simulation scenarios to check the correct implementation of the algorithm. In the performance evaluation scenarios, the target vehicle was driven by four different maneuvers; constant speed with lane keeping, constant speed with lane change, acceleration & deceleration with lane keeping, and acceleration & deceleration with lane change. The maneuvers of the target vehicle in these scenarios can cover the vehicle movement of the urban traffic environment. The performance evaluation is carried out by comparing estimation error of the position and velocity to the reference system. The errors are measured by RTK-GPS, which is precise positioning system, attached on the ego-vehicle and the target vehicle. As a result, the estimation errors of the dynamic states of the track are significantly reduced. The position and velocity errors are reduced up to 76.4% and 51.5%, respectively. Additionally, the target vehicle was continuously tracked without lost while the IbeoLux tracking system re-initializes tracks two or three times in a performance evaluation scenario. |최근 정부의 안전 규제와 고객의 요구사항으로 인해 차량의 안전 및 편의장치의 개발이 급격히 가속화 되고 있다. 이러한 사회적 관심에 따라 여러 가지 안전 및 편의 장치가 상용화 되고 있으며 미래의 지능형 자동차는 궁극적으로 자율주행 자동차로 진화할 것으로 예측되고 있다. 자율주행 자동차는 주행환경을 인지하고 운전자의 개입 없이 스스로 주행하는 차량을 말한다. 이를 실현하기 위해서는 위치추정, 환경인지, 운전계획, 차량제어 및 시스템 운영 등의 기술이 필수적이다. 이 기술들 중에서도 환경인지는 자율주행자동차의 안전한 주행을 위해 위험한 상황을 감지할 수 있는 핵심 기술이라 할 수 있다. 자율주행 자동차는 레이저 스캐너, 레이더, 카메라 등의 다양한 센서를 통해 환경인지를 하게 된다. 이러한 다양한 센서들 중에서도 레이저 스캐너는 고정밀, 고해상도의 거리 정보를 제공하기 때문에 지금까지 개발된 많은 자율주행 자동차들에 사용되고 있다. 레이저 스캐너는 빛을 이용하여 물체와의 거리를 정밀하게 측정하며 회전하는 거울을 이용해 레이저 빔을 반사시켜 2차원 평면을 스캐닝 한다. 레이저 스캐너에 의해 측정된 데이터는 각 방위에 대한 물체의 거리를 나타내므로 극좌표 상의 점들로 나타나게 된다. 검출된 물체를 나타내는 점들의 형태는 레이저 스캐너와의 상대적인 위치에 따라 달라지게 된다. 특히 움직이는 차량에서의 장애물 검출은 그 형태가 시시각각 변하기 때문에 정확한 위치와 움직임의 추정이 어려워진다. 이 논문은 레이저 스캐너에서 검출된 차량의 형상이 시시각각 변함에 따라 나타나는 위치 및 속도 오차를 줄이고자 L-형상 모델 스위칭 기반의 차량추적 알고리즘을 제안한다. 제안한 알고리즘은 L-형상의 코너 점을 이용하여 추적하기 때문에 형상의 변화에 영향을 받지 않는다. 알고리즘은 크게 3가지 부분(L-형상 추출, L-형상 추적, 박스모델 변환)으로 나누어지며 이들은 각각 점 데이터로부터 L-형상을 추출해내고, L-형상의 운동정보와 모양정보를 칼만 필터를 이용하여 추정하며, 마지막으로 차량의 형상에 가까운 박스모델로 정보를 변환하여 정밀한 차량 추적 정보를 얻게 된다. 제안된 차량 추적 알고리즘은 시뮬레이션 환경을 통해 기능이 검증되고, 실제 도로에서 수집된 데이터를 바탕으로 성능이 평가되었다. 알고리즘의 성능은 상업적으로 사용 가능한 레이저 스캐너들 중 가장 성능이 좋다고 알려진 IbeoLux를 기준으로 위치 오차 및 속도 오차를 비교함으로써 평가되었다. 또한 다양한 상황에서 알고리즘의 성능을 분석하기 위해 추적대상 차량은 가∙감속, 차선 유지 및 변경 등의 다양한 주행패턴을 조합하여 운행되었다. 결과적으로 제안된 알고리즘은 IbeoLux에 비해 위치 오차를 최대 76.4%, 속도 오차를 최대 51.5%까지 감소시켰다. 따라서 제안된 알고리즘은 자율주행자동차의 안전한 주행을 위한 인식시스템으로서 우수한 성능을 보임을 증명하였다.; constant speed with lane keeping, constant speed with lane change, acceleration & deceleration with lane keeping, and acceleration & deceleration with lane change. The maneuvers of the target vehicle in these scenarios can cover the vehicle movement of the urban traffic environment. The performance evaluation is carried out by comparing estimation error of the position and velocity to the reference system. The errors are measured by RTK-GPS, which is precise positioning system, attached on the ego-vehicle and the target vehicle. As a result, the estimation errors of the dynamic states of the track are significantly reduced. The position and velocity errors are reduced up to 76.4% and 51.5%, respectively. Additionally, the target vehicle was continuously tracked without lost while the IbeoLux tracking system re-initializes tracks two or three times in a performance evaluation scenario.
URI
https://repository.hanyang.ac.kr/handle/20.500.11754/124673http://hanyang.dcollection.net/common/orgView/200000429517
Appears in Collections:
GRADUATE SCHOOL[S](대학원) > DEPARTMENT OF AUTOMOTIVE ENGINEERING(자동차공학과) > Theses (Ph.D.)
Files in This Item:
There are no files associated with this item.
Export
RIS (EndNote)
XLS (Excel)
XML


qrcode

Items in DSpace are protected by copyright, with all rights reserved, unless otherwise indicated.

BROWSE