서지주요정보
Traffic data imputation and trajectory-level traffic state estimation based on deep learning methods = 딥 러닝 기반 교통정보 결측치 생성 및 궤적 수준 고정밀 교통정보 추정
서명 / 저자 Traffic data imputation and trajectory-level traffic state estimation based on deep learning methods = 딥 러닝 기반 교통정보 결측치 생성 및 궤적 수준 고정밀 교통정보 추정 / Jongho Kim.
발행사항 [대전 : 한국과학기술원, 2024].
Online Access 원문보기 원문인쇄

소장정보

등록번호

8043390

소장위치/청구기호

학술문화관(도서관)2층 학위논문

MGT 24009

휴대폰 전송

도서상태

이용가능(대출불가)

사유안내

반납예정일

리뷰정보

초록정보

This study categorizes traffic detectors into Eulerian and Lagrangian and addresses the challenges each faces. First, Eulerian sensing collects traffic data from all passing vehicles at detector locations over time. However, due to several factors, including detector malfunctions and communication errors, missing data frequently occurs. To address this problem, a Spatial Temporal Transformer Networks (STTNs) model was developed that applies an attention mechanism to impute missing traffic data. This model was evaluated across three types of missing patterns (MCAR, TCM and SCM) and missing rates ranging from 10% to 90%. The evaluation results showed that the accuracy decreases as the missing rate increases, with the highest accuracies observed for MCAR, followed by SCM and TCM. Next, Lagrangian sensing involves collecting trajectory data from probe vehicles, but this data is typically sparse, and detailed trajectory information is obscured during the spatiotemporal aggregation. To address this issue, a Pix2Pix Generative Adversarial Networks (GANs) model was developed, capable of estimating traffic speeds at the trajectory level with only 5% of the probe vehicle data. Despite being trained on simulation data, this model demonstrated superior performance on real-world datasets from NGSIM US-101 and I-80. These results suggest that even with varying road geometries and traffic characteristics, traffic states can be accurately estimated at the trajectory level based on images derived from trajectory data.

본 연구는 교통 검지기 유형을 오일러리안과 라그랑지안 검지기로 분류하고, 각 검지 방식이 직면하는 문제를 해결하고자 한다. 먼저, 오일러리안 검지기는 시간 경과에 따라 검지기가 설치된 지점을 통과하는 모든 차량의 데이터를 수집한다. 그러나, 검지기의 오작동이나 통신 오류와 같은 다양한 요인으로 인해 데이터 누락이 종종 발생한다. 이를 해결하고자, 어텐션 메커니즘을 적용한 Spatial Temporal Transformer Networks (STTNs) 모델을 개발하여 누락된 교통 데이터를 재구성하였다. 이 모델은 세가지 유형의 누락 패턴 (MCAR, TCM, SCM)과 10%에서 90%에 이르는 다양한 누락률에서 평가되었다. 평가 결과, 누락률이 증가함에 따라 정확도가 낮아 졌으며, MCAR, SCM, TCM 순으로 높은 정확도를 보였다. 다음, 라그랑지안 검지기는 프로브 차량으로부터 궤적 데이터를 수집하나 이 데이터는 표본이 매우 희박하며, 시공간 집계 과정에서 궤적의 세부 정보가 손상되는 문제가 있다. 이를 해결하고자, 단 5% 프로브 차량 데이터만을 활용하여 궤적 수준에서 교통 속도를 추정할 수 있는 Pix2Pix Generative Adversarial Networks (GANs) 모델을 개발하였다. 이 모델은 시뮬레이션 데이터를 기반으로 학습되었음에도, 실 데이터인 NGSIM US-101과 I-80에서 우수한 성능을 보였다. 이 결과는 도로의 기하학적 구조와 교통 특성이 다름에도 불구하고, 5% 궤적 데이터를 기반으로 형상화된 이미지 레이아웃만으로도 궤적 수준에서 교통 상태를 정확하게 추정할 수 있음을 보여준다.

서지기타정보

서지기타정보
청구기호 {MGT 24009
형태사항 v, 67 p. : 삽도 ; 30 cm
언어 영어
일반주기 저자명의 한글표기 : 김종호
지도교수의 영문표기 : Kitae Jang
지도교수의 한글표기 : 장기태
Including appendix
학위논문 학위논문(석사) - 한국과학기술원 : 조천식모빌리티대학원,
서지주기 References : p. 60-64
주제 Traffic Data Imputation
Traffic State Estimation
Eulerian Sensing
Lagrangian Sensing
Trajectory Data
Deep Learning
결측치 생성
교통정보 추정
오일러리안 검지기
라그랑지안 검지기
궤적데이터
딥 러닝
QR CODE

책소개

전체보기

목차

전체보기

이 주제의 인기대출도서

Traffic Sensing Approaches

Traffic Missing Patterns: (a) MCAR; (b) TCM; (c) SCM

Aggregation of Trajectory-level Traffic State based on Sparse Probe Vehicles

Framework of the Thesis

Traffic State Estimation Methods Source: Seo et al., 4

Data Sites: (a) I-80 WB; (b) US-101 SB

Temporal dependency: Example of traffic flow at VDS ID 408101 from May 1, 2023, to May 7, 2023

Example ofthe contour map on I-80 WB, recorded by PeMS D4 on Wed., May 3, 2023: (a) Speed; (b) Traffic Flow

Tensor Dimensions of Training and Validation Datasets

Spatial Temporal Transformer Networks Architecture

Self Attention Mechanism: (a) Spatial Transformer Architecture; (b) Temporal Transformer Architecture

Loss of the Proposed Model: (a) I-80 WB Speed; (b) I-80 WB Traffic Flow; (c) US-101 SB Speed; (d) US-101 SB Traffic Flow

Example ofImputed Speed by MCAR pattern at VDS ID 400060 from Oct. 30, 2023, to Nov 5, 2023: (a) Observation; (b) Imputed Speed; (c) Residual (mi/h)

Example of Imputed Traffic flow by MCAR pattern at VDS ID 400060 from Oct. 30,2023 to Nov. 5. 2023: (a) Observation; (b) Imputed Traffic flow; (c) Residual (veh/5-minute)

Imputation Performance for MCAR Pattern

Example of Imputed Speed by TCM pattern at VDS ID 400060 from Oct. 30, 2023, to Nov. 5. 2023: (a) Observation; (b) Imputed Speed; (c) Residual (mi/h)

Example of Imputed Traffic flow by TCM pattern at VDS ID 400060 from Oct. 30, 2023 to Nov. 5, 2023: (a) Observation: (b) Imputed Traffic flow; (c) Residual (veh/5-minute)

Imputation Performance for TCM Pattern

Example of Imputed Speed by SCM pattern at VDS ID 400060 from Oct. 30, 2023, to Nov. 5, 2023: (a) Observation; (b) Imputed Speed; (c) Residual (mi/h)

Example of Imputed Traffic flow by SCM pattern at VDS ID 400060 from Oct. 30, 2023, to Nov. 5, 2023: (a) Observation; (b) Imputed Traffic flow; (c) Residual (veh/5-minute)

Imputation Performance for SCM Pattern

Performance Evaluation of Speed Imputation on I-80 WB and US-101 SB

Performance Evaluation of traffic flow Imputation on I-80 WB and US-101 SB

Example of Attention Score According to MCAR Pattern on Oct. 24, 2023, for US-101 SB: (a) MACR Pattern; (b) Attention Score of Spatial Transformer at 08:00; (c) Attention Score of Temporal Transformer at Station ID 400819

Example of Attention Score According to TCM Pattern on Oct. 24, 2023, for US-101 SB: (a) TCM Pattern; (b) Attention Score of Spatial Transformer at 08:00; (c) Attention Score of Temporal Transformer at Station ID 400819

Example of Attention Score According to SCM Pattern on Oct. 24, 2023, for US-101 SB: (a) SCM Pattern; (b) Attention Score of Spatial Transformer at 08:00; (c) Attention Score of Temporal Transformer at Station ID 400819

NGSIM geometry: (a) US-101; (b) I-80

Example Trajectory Data for Free-Flow, Transition, and Congestion States from Simulatior

Definitions of Variables

Microscopic traffic state: 3-dimension tensor

Pix2Pix GAN Structure

Generator Loss

Number of Vehicle Trajectories per Sparse Image in the Test Dataset

Selected examples from sparse image to Reconstruction of the Simulation dataset

Selected examples from sparse image to Reconstruction of the NGSIM US-101 dataset

Selected examples from sparse image to Reconstruction of the NGSIM I-80 dataset

Histogram showing PSNR and SSIM between the actual and reconstructed images based on Simulation and NGSIM dataset

Statistical Summary of SSIM and PSNR for Different Datasets

Statistical Summary of RMSE and MAE between the actual and estimated speed

Histogram showing RMSE and MAE between the actual and estimated speed based on Simulation and NGSIM dataset

Example ofselected TSE using Space-Time Diagram in NGSIM US-101. (c) with RMSE of 8.693 km/h and MAE of6.750 km/h. (d) with RMSE of8.440 km/h and MAE of6.130 km/h

Results of proposed TSE in NGSIM US-101

Example of selected TSE using Space-Time Diagram in NGSIM I-80: (c) with RMSE of 9.949 km/h and MAE of 7.646 km/h; (d) with RMSE of7.661 km/h and MAE of5.151 km/h

Results of proposed TSE in NGSIM I-80

Flow-Density Scatter Plot of Traffic Characteristics Derived from Image

Examples ofselected inaccurate image reconstructions: (a) Example ofleft-skew oftrajectory with RMSE of17.757 km/h and MAE of11.915 km/h; (b) Example ofright-skew oftrajectory with RMSE of 26.271 km/h and MAE of 19.618 km/h; (c) Example of an extremely sparse trajectory with RMSE of 25.392 km/h and MAE of 20.976km/h