Indoor Parking Localization based on Dual Weighted Particle Filter
- 주제(키워드) 위치인식 , 자율주차
- 발행기관 고려대학교 대학원
- 지도교수 홍대희
- 발행년도 2017
- 학위수여년월 2017. 2
- 학위구분 석사
- 학과 대학원 자동차융합학과
- 원문페이지 57 p
- 실제URI http://www.dcollection.net/handler/korea/000000073258
- 본문언어 영어
- 제출원본 000045897518
초록/요약
The autonomous vehicle is capable of sensing its environment and navigating without human control input. For the fully autonomous vehicle, the autonomous valet parking technology is required. The most important technology of the autonomous valet parking is knowing its location in the global map. In the outside car park, the GPS transmits a position data to the vehicle, but it is not possible in an indoor parking lot. Localization in the indoor parking lot faces two challenges. The first one is the absence of GPS coverage. For an autonomous vehicle driving inside a parking lot safely, the precise recognition about location of the autonomous vehicle is essential. The alternative method to replace GPS coverage is using a localization algorithm with data from several sensors and a map of environment. Second, the environment in the parking space is not static. There are vehicles in parking slots and their environment always changes over time. Generally, the map of parking lot does not contain information of the parking space usage. This map contains only static environment of the parking lot. It results in limitation of the raw data map matching. It means that autonomous vehicle cannot obtain the same environment data from sensor that has to match up with map data. We can handle this problem with a feature extraction algorithm. It is assumed that the map has data about pillar and parking slot’s positions. With a sensor data processing, the position of parking slots and pillars can be estimated. This paper proposes a robust algorithm to localize in the non-static nature of parking area without GPS. This algorithm is developed to take the advantage of using map of parking lot to get necessary pose accuracy. There are other solutions to achieve localization without map data. Localization with SLAM algorithm is the one of them. But it is not feasible in the parking lot due to the dynamic nature of the parking vehicles. Another one is using only odometry data. It is also not appropriate because of the biased data and large error. For this reason, it is necessary to take advantage of using map data for the robust localization algorithm even though a parking place is a hard condition to apply localization algorithm with a map. The map used in this algorithm consists of pillars and parking slots. Their center position are extracted to match up with map. Pillars were considered as static objects. Parked vehicles are considered semi-static objects. Because parked vehicles change every situation. Position error occur during matching the vehicle position center and parking slot’s center for using the map feature data. It is reduced using pillar observation data. This algorithm is based on the particle filter. This algorithm contains map features that have not only positions but also covariance information. Different covariance are assigned to parking slots and pillars. Particle filter has a resampling step to minimize errors. These errors can be reduced when pillar data is obtained. Because there are no alignment errors with it. The algorithm also can be used in a harsh condition that pillar is not visible. The algorithm was evaluated in the scaled down indoor parking model with five environments scenario. Scenario 1 is an empty vehicle situation. This experiment was performed to compare the algorithm’s performance with basic particle filter. Because there are not vehicles with error position, the proposed algorithm’s result is expected to the same value with basic particle filter. Scenario 2 is an eight and all parked vehicle with no alignment error situation. This experiment was performed to validate the feature extraction algorithm and data association problem. The scenario 3 is a five vehicles with alignment error. This experiment was organized to verify location accuracy. Centers of five vehicles were not matched with the center positions of parking slots and it results in the performance difference between basic particle filter and dual weighted particle filter. The scenario 4 is the all vehicles parked with error. It is hard condition to extract pillar position. We did this experiment to verify total algorithm’s performance and robustness. The evaluation of the algorithm was based on the motion analysis system. The feature extraction algorithm and data association was done accurately. Our algorithm had higher position accuracy than basic particle filter in all case except scenario 1. The maximum RMS absolute position error was 12cm.
more목차
Contents
Abstract………………………………………………………………………………i
Contents…………………………………………………………………………… v
List of Figures………………………………………………………………………vi
List of Tables………………………………………………………………………….x
CHAPTER 1 1
1.1 RESEARCH BACKGROUND 1
1.2 RESEARCH PURPOSE 4
1.3 THESIS OVERVIEW 4
CHAPTER 2 6
2.1 OUTLINE OF PARTICLE FILTER 6
2.2 MOTION MODEL 8
2.3 MEASUREMENT MODEL 9
2.4 RESAMPLING ALGORITHM 10
2.5 DATA ASSOCIATION 11
CHAPTER 3 13
3.1 PROBLEM DEFINITION 13
3.2 IMPORTANCE WEIGHT EQUATION 15
3.3 DUAL WEIGHTING 16
3.4 COVARIANCE CALCULATION 19
CHAPTER 4 21
4.1 DOUGLAS PEUCKER ALGORITHM 21
4.2 FEATURE EXTRACTION METHOD 22
4.2.1 Clustering 22
4.2.2 Segmentation 23
4.3.3 Feature point extraction 24
4.3.4 Feature extraction experiments 26
CHAPTER 5 28
5.1 EXPERIMENTAL CONFIGURATION 28
5.2 EXPERIMENTAL SCENARIO 31
5.3 EXPERIMENTAL RESULTS 34
CHAPTER 6 39
REFERENCES 41
