S-Space College of Engineering/Engineering Practice School (공과대학/대학원) Dept. of Mechanical Aerospace Engineering (기계항공공학부) Theses (Master's Degree_기계항공공학부)
Surrounding Vehicle Perception Algorithm in Intersection for Autonomous Vehicle with Laser scanner
레이저 스캐너 기반 자율주행용 교차로 내 주변 차량 인지 알고리즘 개발
- 공과대학 기계항공공학부
- Issue Date
- 서울대학교 대학원
- Vehicle Perception Filter; Point Cloud Post-processing; Particle Filter; Autonomous Vehicle; Laser Scanner; Recursive Structure
- 학위논문 (석사)-- 서울대학교 대학원 공과대학 기계항공공학부, 2017. 8. 이경수.
- The aim of this study is designing surrounding vehicle movement perception algorithm in urban condition. In recent autonomous vehicle industry, many researchers are focusing on three major topics which is called environment perception, localization and designing controller for autonomous vehicle these days. Especially for the perception technology, the design of the algorithm follows the characteristics of the target environment or objects. In Urban condition, to design safe drivable path for autonomous vehicle, the objects’ position is much important than high way conditions. Also, various objects appear which cannot be detected with RADAR or yield fault case with camera. Because of these reasons, many research are trying to fuse different sensors including camera, RADAR and LiDAR to overcome the challenges that can occur in urban conditions, therefore, laser scanner based target detecting technology is needed to perceive in city road.
The tracking filter consists of two parts, shape estimation and tracking filter. To fuse with other sensors or designing target filter, there should be a step for compressing point cloud group information into some representative point or state. Thus in shape estimation parts, we transform the laser scanner’s point cloud data into vehicle position state measurement value. Vehicle shape estimation also consists of two parts, clustering and shape extraction. Clustering classify the total point cloud into object level and shape extraction estimates the vehicle liked objects’ position information. The clustering part works based on Euclidean Minimum Spanning Tree (EMST), and for the shape extraction, Random Sample Consensus (RANSAC) method is used to estimate the target objects rear and side edge. The second part, tracking filter, has two different filters. Particle filter estimates the target vehicle’s position including heading angle. To improve the tracking performance of the particle filter, Kalman filter is also designed to estimate the velocity and yaw rate recursively to update the process model of the particle filter.
The performance of the proposed algorithm has been verified with several stages. To check quantitative error level, off-line simulation is held for profile based motion tracking case and designed intersection simulator with simple path tracking algorithm for the target vehicle. In these conditions, the exact target vehicle’s position information was known, thus we verified the error level of the lateral/longitudinal direction of target vehicle’s local coordinate which is important information when designing driving path or controller. For the second step, simulation with point cloud data which is collected from the test vehicle was held to verify its performance for actual vehicle condition. As a final stage, for integrating into autonomous vehicle, the proposed algorithm evaluated into the test vehicle for guaranteeing on-line performance.