SUMO-enhanced traffic anomaly detection: enhancing spatiotemporal transferability through simulation-based data augmentation

Haonan Yang; Chenlei Liao; Zhiwu Dong; Chuqiao Chen; Dong Mo; Wenhao Chen; Haonan Yang; Chenlei Liao; Zhiwu Dong; Chuqiao Chen; Dong Mo; Wenhao Chen

doi:10.48130/dts-0025-0025

Traffic anomaly detection is crucial for road safety and traffic efficiency, but the scarcity of accident data poses challenges for model training. Although current research has utilized deep neural networks for data augmentation in small-sample scenarios, it has largely overlooked real-world physical constraints and lacks exploration into zero-shot scenarios, areas that still require further development. The paper proposes a SUMO-based simulation-driven data augmentation approach that enhances detection performance under limited real-world data conditions. Synthetic accident data are generated under five physical constraints and integrated with real traffic data to augment the training set. This study evaluates the method using three deep learning models (MLP, CNN, and LSTM) on PeMS datasets from two California highways (I80-E and I10-E). Experimental results demonstrate that the simulation-augmented training significantly improves model performance. In the spatial transfer experiment, the MLP model's accuracy improved from 75.53% to 76.58%, and AUC increased from 53.62% to 56.74%. These findings indicate that this approach effectively enhances both temporal and spatial transferability for traffic anomaly detection in small-sample scenarios.

HTML

Introduction

Highways, as a critical component of modern transportation systems, have significantly promoted interconnectivity between regions and provided a solid foundation for national economic development. However, traffic accidents remain one of the primary public safety challenges worldwide, causing substantial fatalities, injuries, and economic losses annually. According to the World Health Organization (WHO), traffic accidents are the leading cause of death for individuals aged five to 29, accounting for approximately 1.19 million fatalities per year^[1]. Additionally, millions of injuries resulting from these accidents impose a heavy burden on society. With the acceleration of urbanization and the increasing number of vehicles, the issue of frequent traffic accidents has become more pronounced, prompting academia and industry to urgently develop efficient technologies for rapid accident detection and response.

Currently, various accident prediction and detection techniques have been developed to improve traffic safety through risk identification. These techniques can be categorized into two types: statistical models and machine learning-based models^[2]. Traditional statistical models rely on historical data and mathematical analysis to uncover patterns within the data. However, they exhibit limitations when addressing complex scenarios and large-scale datasets^[3]. In contrast, with the rapid development of artificial intelligence (AI) technology^[4,5], AI-based models have shown greater potential in traffic accident prediction and detection. These methods extract complex features from large-scale data and achieve high-precision predictions. Nevertheless, AI-based solutions are highly dependent on high-quality data, and the traffic accident data often suffers from insufficient quantity and quality. This issue limits the effectiveness of AI technologies in practical applications. Advanced data cleaning and preprocessing methods can address this challenge by improving data quality and enriching traffic event datasets with additional information.

Despite increasing interest in using deep learning for traffic accident prediction and detection, significant limitations remain. Most existing data augmentation strategies—such as those using Generative Adversarial Networks (GANs)—focus on statistical realism without incorporating real-world physical constraints. These approaches may produce synthetic data that lacks behavioral interpretability or alignment with actual traffic dynamics. Meanwhile, advanced traffic simulation platforms like SUMO have been extensively used for traffic modeling and control evaluation, but their potential as tools for generating physically grounded synthetic data to improve anomaly detection models remains underexplored. Furthermore, very few studies investigate the ability of augmented models to generalize in zero-shot or data-scarce scenarios, particularly across different temporal or spatial domains.

To address the aforementioned challenges, the paper proposes a SUMO (Simulation of Urban Mobility) -based simulation data augmentation method aimed at improving the performance of traffic accident detection and prediction by enhancing data quality and coverage. Specifically, the main contributions of the paper are as follows:

(1) Incorporation of physical constraints: by integrating traffic-related physical information, the paper introduces a physical-informed model that incorporates physical constraints into the data augmentation process, addressing the lack of physical constraints in existing methods. These physical constraints include maximum acceleration, comfortable deceleration, minimum safe distance, acceleration index, and driver reaction time. While these five parameters capture key aspects of vehicle dynamics and are directly supported in SUMO, other influential real-world factors, such as weather conditions or driver behavior variability, were not included due to limited simulator support and the scope of this study. Future work will explore incorporating such factors to further enhance the realism of the augmented data.

(2) Temporal and spatial transfer modeling: using the SUMO software, the study conducts temporal transfer modeling for roads with limited accident data and spatial transfer modeling for roads without accident data, thereby extending the temporal and spatial applicability of the data.

(3) Deep learning-based accident detection: the study utilized three commonly applied neural network models for accident detection and validated the model's detection performance using real-world data and simulation data. Experimental results demonstrate that the model significantly improves detection performance and reliability.

The remainder of this paper is organized as follows: the 'related work' section reviews existing studies on traffic accident detection and prediction, and surveys relevant research on traffic flow data augmentation. The 'methodology' section introduces the SUMO simulation techniques and the accident prediction and detection models, including multilayer perceptron (MLP), convolutional neural network (CNN), and long short-term memory (LSTM). The 'data preparation' section details the data sources, selected roads, and data preprocessing methods. The 'results' section presents the simulation-generated normal traffic flows, temporal and spatial transfer modeling, and experimental results. Finally, the 'discussion and contributions' section summarizes the main contributions of the paper and highlights potential directions for future research.

Related work

Accident prediction and detection have long been central topics in transportation research. With the rapid advancement of machine learning (ML), AI-based models have gradually become a core area of research, leading to numerous studies that utilize traffic accident data for analysis, prediction, and detection.

Accident prediction
Research in accident prediction primarily falls into two categories: statistical analysis models and machine learning-based models. Statistical analysis, such as logistic regression^[6−8], log-linear models^[9], Bayesian logistic models^[10], and Bayesian dynamic logistic regression models^[11], provide valuable insights, however, they often struggle to capture the complex nonlinear relationships inherent in traffic flow data, leading to limited performance.

In contrast, ML models demonstrate greater potential due to their ability to handle high-dimensional and nonlinear data. Huang et al.^[12] utilized deep learning techniques to predict crash risks, employing data from radar sensors on Interstate 235 in Des Moines, IA. Their model demonstrated notable improvements in detection accuracy. Pourroostaei et al.^[13] explored road car accident data patterns and proposed a data preprocessing model, including missing data removal, attribute generalization, and outlier removal, to improve the accuracy of traffic accident prediction using machine learning techniques.

There are likewise many other studies that have investigated accident risk prediction by fusing traffic data and combining statistical and ML methods. For instance, Yu et al.^[14] proposed a Deep Spatio-Temporal Graph Convolutional Network (DSTGCN) to predict traffic accidents by effectively learning spatial, temporal, and external data correlations, and achieved superior performance on real-world datasets. Guo et al.^[15] incorporated risky driving behavior and traffic flow data to predict traffic accident risks, achieving a prediction accuracy of 84.48%. Zhang et al.^[16] developed a bidirectional LSTM model with convolutional layers to analyze roadside detector data and connected vehicle information, reaching a recall of 0.772 and an AUC of 0.857. Yuan et al.^[17] combined statistical and ML models to explore the relationship between traffic conflicts and flow features using the HighD dataset, highlighting the significant influence of traffic flow on conflict probability.

Accident detection
Early-stage methods for incident detection predominantly relied on comparative algorithms, such as the well-known California algorithm^[18], which compares traffic parameters (e.g., speed, occupancy) to predefined thresholds^[19]. While effective, these methods depend heavily on fixed detector readings and lack adaptability.

With the development of computational intelligence, Artificial Intelligence algorithms apply AI and ML models to identify the normal and abnormal traffic patterns and then classify the given input data as either incident or normal conditions^[20]. For example, Dogru et al.^[21] proposed an intelligent traffic accident detection system based on VANETs and machine learning, using vehicle speed and position data for high-accuracy accident detection with the Random Forest algorithm. White et al.^[22] proposed a smartphone-based traffic accident detection system that combines accelerometers, acoustic data, and context information to automatically notify emergency services, provide situational awareness, and reduce false positives. Ozbayoglu et al.^[23] designed a real-time autonomous detection system using computational intelligence techniques to analyze Istanbul's traffic data, predicting accidents with nearest neighbor models and feed-forward neural networks. Gu et al.^[24] explored social media as an alternative data source, mining Twitter for real-time traffic incident detection and classification. Mehrannia et al.^[25] applied LSTM networks for freeway accident detection, improving class separability and reducing dimensionality.

Data augmentation
The above AI-based solutions for incident detection have highlighted challenges, particularly regarding data quantity and quality. Traditional traffic datasets often lack sufficient volume and exhibit quality issues, limiting their applicability. To address these challenges, data enhancement techniques have emerged as a critical research focus.

Lin et al.^[26] introduced a hybrid model using Generative Adversarial Networks (GANs) to expand sample size and balance datasets, complemented by Temporal and Spatially Stacked Autoencoders (TSSAE) for feature extraction. Xie et al.^[27] proposed SASYNO-RF-RSKNN, a hybrid method tackling imbalanced datasets using oversampling techniques and ensemble algorithms. Similarly, Jilani et al.^[28] utilized GAN-based augmentation to train CNN models, achieving high classification accuracy. Other studies emphasize advanced augmentation strategies. Huang et al.^[29] integrated GANs with TSSAE to address sample limitations and detect temporal and spatial traffic patterns. Dabboussi et al.^[30] employed GANs to generate realistic traffic speed data, improving prediction for specific scenarios. Benabdallah Benarmas et al.^[31] highlighted the role of Deep Generative Models (DGMs) in enhancing traffic speed predictions through synthetic dataset generation.

Summarily, these advancements underscore the significant progress in accident prediction and detection, particularly through the integration of deep learning and data augmentation. However, challenges remain in addressing the dependency on high-quality data and balancing the scalability of models for diverse traffic scenarios are key. Based on these insights, this paper proposes a method that combines SUMO simulation data with real-world data. This method not only increases the diversity of the data but also effectively improves the performance of the model.

Methodology

To enhance the performance of traffic accident detection systems, the SUMO was utilized for data augmentation and advanced deep learning techniques for traffic accident detection. The following section first introduces SUMO, followed by an introduction to the deep learning models used.

Introduction to SUMO
SUMO is a microscopic traffic simulation platform designed for modeling multi-modal transportation systems, including vehicles, public transit, pedestrians, and cyclists, developed by the Dresden University of Technology in Germany^[37]. Since its inception in 2001, the SUMO software suite has been widely applied in numerous domestic and international research projects. Its primary applications include traffic signal evaluation^[38], route selection and re-planning^[39], traffic monitoring method assessment^[40], vehicle communication simulation^[41], and traffic forecasting^[42]. As an open-source software, SUMO allows users to customize and extend it according to their needs. It supports multiple traffic modes, such as cars, pedestrians, and public transport, and can be integrated with tools like Python and MATLAB, facilitating data analysis and model training. Its high efficiency enables SUMO to quickly process large volumes of traffic data and simulate complex traffic scenarios, making it particularly suitable for tasks such as traffic anomaly detection and model optimization. Additionally, SUMO possesses powerful data generation and enhancement capabilities, enabling the creation of simulation data with complex traffic patterns and diverse scenarios, which are especially well-suited for training deep learning models. Finally, SUMO simulates based on real-world traffic network data and employs precise vehicle behavior modeling, ensuring that the simulation results closely align with real-world traffic conditions, demonstrating applicability in real-world scenarios. Due to the various advantages of SUMO mentioned above, SUMO was ultimately chosen as the simulation environment in this paper to generate traffic accident augmented data.

Simulation calibration

The Intelligent Driver Model (IDM), a microscopic traffic simulation model designed to replicate vehicle dynamics and driver behavior, was calibrated using an exhaustive grid search method. IDM employs precise mathematical formulations to simulate how vehicles adjust their speed and spacing in response to leading vehicles under various traffic conditions, considering factors such as desired speed, current speed, gap distance, and traffic density to determine optimal acceleration or deceleration strategies.

The IDM model assumes that the driver will dynamically adjust the acceleration according to the behavior of the vehicle in front and the current traffic conditions. The acceleration of the vehicle is determined by the following factors: (1) Expected free travel speed: When there is no interference from the vehicle in front, the vehicle tends to travel at a desired speed (such as the speed limit); (2) Safety distance: The vehicle will adjust its speed according to the distance and relative speed to the vehicle in front to avoid a rear-end collision; and (3) Acceleration and deceleration capabilities: The physical properties of the vehicle also constrain driving behavior.

The acceleration formula of the IDM model is as follows:

$ \dot{v}=a\left[1-{\left(\dfrac{v}{{v}_{0}}\right)}^{\delta }-{\left(\dfrac{{s}^{*}\left(v,\Delta v\right)}{s}\right)}^{2}\right] $

(1)

where, a represents the maximum acceleration, v represents the current vehicle speed, v₀ represents the expected speed, δ represents the acceleration index. Specifically, δ controls how aggressively or smoothly a vehicle approaches its desired speed. A larger value of δ leads to smoother, more gradual acceleration, promoting comfort and stability, while a smaller value results in faster, more abrupt acceleration responses. s represents the distance between the current vehicle and the vehicle in front, and s*(v, Δv) represents the expected distance, which can be expressed as:

$ {s}^{*}\left(v,\Delta v\right)={s}_{0}+max\left(0,vT+\dfrac{v\Delta v}{2\sqrt{ab}}\right) $

(2)

where, Δv represents the relative speed, that is, the speed difference between the current vehicle and the vehicle in front; s₀ represents the minimum safe distance, b represents the comfortable deceleration. If the current vehicle suddenly decelerates, the vehicle will adjust the current speed with the deceleration of b to avoid a rear-end collision; T represents time headway, which refers to the driver's reaction time.

Parameter calibration is used to determine the IDM parameters. The objective of calibration was to minimize the Root Mean Square Error (RMSE) between simulated traffic flow and real-world detector readings. The calibration process was conducted over a defined search space constrained by physically reasonable ranges for each IDM parameter, including maximum acceleration and comfortable deceleration. The upstream section of the three-lane road section I80-E, which is monitored by four detectors (402241, 402243, 402245, 402246) was selected. The vehicle data from the upstream detectors is used as input, and the difference between the simulated and actual readings from the downstream detectors is used to fine-tune the IDM parameters to ensure that the simulation accurately reflects the observed traffic behavior.

Vehicle inject

After adopting the calibrated parameters, let the probability of a vehicle appearing per second be p. Then, the number of vehicles departing on this lane within 5 min, X, follows a binomial distribution:

$ X \sim B\left(300,p\right) $ (3)

The mean of this distribution can be expressed as E(X) = 300p, which equals the actual number of vehicles N detected by the detector during the 5 mins. Therefore, the input probability p = N/300.

Accident detection methods
MLP is a classic feedforward neural network architecture composed of multiple layers, typically including an input layer, one or more hidden layers, and an output layer. Each neuron in one layer is fully connected to the neurons in the previous layer, processing data through weighted sums, biases, and activation functions to transmit information through the network. The goal of MLP is to accurately model the complex nonlinear relationships between input data and outputs, enabling precise classification or regression tasks.

CNNs are a class of deep learning models primarily used for processing and analyzing grid-like data structures, such as images or time-series data. CNNs excel at automatically learning spatial hierarchies of features, making them particularly effective in tasks like image classification, object detection, and anomaly detection.

LSTM networks are a specialized type of Recurrent Neural Network (RNN) designed to handle and predict long-term dependencies in sequential data. Traditional RNNs often struggle with long sequences due to the vanishing or exploding gradient problem, making it difficult to effectively retain information from distant points in the sequence. LSTM addresses this issue by introducing 'memory cells' and 'gating mechanisms,' which enable the model to manage information flow more effectively.

This paper implements the prediction of normal traffic flow using LSTM to validate the feasibility of using SUMO simulation data, and accident detection of abnormal traffic flow using MLP, CNN, and LSTM, using multiple neural networks with the aim of eliminating chance. These models are chosen for their simplicity, interpretability, and compatibility with small-sample datasets. While more advanced models such as spatiotemporal graph convolutional networks (STGCNs) or graph neural networks (GNNs) can more effectively capture complex spatiotemporal dependencies, they usually require larger-scale graph-structured data and more extensive training samples. Given the small-sample and data-scarce scenarios emphasized in this work, simpler models provide more robust and interpretable benchmarks to evaluate the effectiveness of simulation-based data augmentation. Moreover, the consistent improvements demonstrated in these commonly used models reinforce the practical value of our proposed approach.

[1]	World Health Organization. 2023. Road traffic injuries. www.who.int/zh/news-room/fact-sheets/detail/road-traffic-injuries (Retrieved December 21, 2024)
[2]	Ghosh B, Smith DP. 2014. Customization of automatic incident detection algorithms for signalized urban arterials. Journal of Intelligent Transportation Systems 18(4):426−41 doi: 10.1080/15472450.2013.806843 CrossRef Google Scholar
[3]	Li L, Qin L, Qu X, Zhang J, Wang Y, et al. 2019. Day-ahead traffic flow forecasting based on a deep belief network optimized by the multi-objective particle swarm algorithm. Knowledge-Based Systems 172:1−14 doi: 10.1016/j.knosys.2019.01.015 CrossRef Google Scholar
[4]	Al Kuwaiti A, Nazer K, Al-Reedy A, Al-Shehri S, Al-Muhanna A, et al. 2023. A review of the role of artificial intelligence in healthcare. Journal of Personalized Medicine 13(6):951 doi: 10.3390/jpm13060951 CrossRef Google Scholar
[5]	Wang H, Fu T, Du Y, Gao W, Huang K, et al. 2023. Scientific discovery in the age of artificial intelligence. Nature 620:47−60 doi: 10.1038/s41586-023-06221-2 CrossRef Google Scholar
[6]	Abdel-Aty M, Uddin N, Pande A, Abdalla MF, Hsia L. 2004. Predicting freeway crashes from loop detector data by matched case-control logistic regression. Transportation Research Record 1897:88−95 doi: 10.3141/1897-12 CrossRef Google Scholar
[7]	Abdel-Aty M, Uddin N, Pande A. 2005. Split models for predicting multivehicle crashes during high-speed and low-speed operating conditions on freeways. Transportation Research Record 1908(1):51−58 doi: 10.1177/0361198105190800107 CrossRef Google Scholar
[8]	Oh C, Oh JS, Ritchie SG. 2005. Real-time hazardous traffic condition warning system: Framework and evaluation. IEEE Transactions on Intelligent Transportation Systems 6(3):265−72 doi: 10.1109/TITS.2005.853693 CrossRef Google Scholar
[9]	Lee C, Abdel-Aty M. 2006. Temporal variations in traffic flow and ramp-related crash risk. In Applications of Advanced Technology in Transportation. Chicago: American Society of Civil Engineers. pp. 244–49 doi: 10.1061/40799(213)40
[10]	Yu R, Abdel-Aty M. 2013. Multi-level Bayesian analyses for single-and multi-vehicle freeway crashes. Accident Analysis & Prevention 58:97−105 doi: 10.1016/j.aap.2013.04.025 CrossRef Google Scholar
[11]	Yang K, Wang X, Yu R. 2018. A Bayesian dynamic updating approach for urban expressway real-time crash risk evaluation. Transportation Research Part C: Emerging Technologies 96:192−207 doi: 10.1016/j.trc.2018.09.020 CrossRef Google Scholar
[12]	Huang T, Wang S, Sharma A. 2020. Highway crash detection and risk estimation using deep learning. Accident Analysis & Prevention 135:105392 doi: 10.1016/j.aap.2019.105392 CrossRef Google Scholar
[13]	Pourroostaei Ardakani S, Liang X, Mengistu KT, So RS, Wei X, et al. 2023. Road car accident prediction using a machine-learning-enabled data analysis. Sustainability 15(7):5939 doi: 10.3390/su15075939 CrossRef Google Scholar
[14]	Yu L, Du B, Hu X, Sun L, Han L, et al. 2021. Deep spatio-temporal graph convolutional network for traffic accident prediction. Neurocomputing 423:135−47 doi: 10.1016/j.neucom.2020.09.043 CrossRef Google Scholar
[15]	Guo M, Zhao X, Yao Y, Yan P, Su Y, et al. 2021. A study of freeway crash risk prediction and interpretation based on risky driving behavior and traffic flow data. Accident Analysis & Prevention 160:106328 doi: 10.1016/j.aap.2021.106328 CrossRef Google Scholar
[16]	Zhang S, Abdel-Aty M. 2022. Real-time crash potential prediction on freeways using connected vehicle data. Analytic Methods in Accident Research 36:100239 doi: 10.1016/j.amar.2022.100239 CrossRef Google Scholar
[17]	Yuan C, Li Y, Huang H, Wang S, Sun Z, et al. 2022. Using traffic flow characteristics to predict real-time conflict risk: a novel method for trajectory data analysis. Analytic Methods in Accident Research 35:100217 doi: 10.1016/j.amar.2022.100217 CrossRef Google Scholar
[18]	Payne HJ, Tignor SC. 1978. Freeway incident-detection algorithms based on decision trees with states. Transportation Research Record 1978(682):30−37 Google Scholar
[19]	Parkany E, Xie C. 2005. A complete review of incident detection algorithms & their deployment: what works and what doesn't. Technical Report NETCR37. New England Transportation Consortium, USA. https://onlinepubs.trb.org/onlinepubs/trispdfs/00988875.pdf
[20]	Jin X, Srinivasan D, Cheu RL. 2001. Classification of freeway traffic patterns for incident detection using constructive probabilistic neural networks. IEEE Transactions on Neural Networks 12(5):1173−87 doi: 10.1109/72.950145 CrossRef Google Scholar
[21]	Dogru N, Subasi A. 2018. Traffic accident detection using random forest classifier. 2018 15^th Learning and Technology Conference (L&T), Jeddah, Saudi Arabia, 25–26 February 2018. USA: IEEE. pp. 40–45 doi: 10.1109/LT.2018.8368509
[22]	White J, Thompson C, Turner H, Dougherty B, Schmidt DC. 2011. WreckWatch: Automatic traffic accident detection and notification with smartphones. Mobile Networks and Applications 16:285−303 doi: 10.1007/s11036-011-0304-8 CrossRef Google Scholar
[23]	Ozbayoglu M, Kucukayan G, Dogdu E. 2016. A real-time autonomous highway accident detection model based on big data processing and computational intelligence. 2016 IEEE International Conference on Big Data (Big Data), Washington, DC, USA, 5–8 December 2016. USA: IEEE. pp. 1807–13 doi: 10.1109/BigData.2016.7840798
[24]	Gu Y, Qian Z, Chen F. 2016. From Twitter to detector: Real-time traffic incident detection using social media data. Transportation Research Part C: Emerging Technologies 67:321−42 doi: 10.1016/j.trc.2016.02.011 CrossRef Google Scholar
[25]	Mehrannia P, Bagi SSG, Moshiri B, Al‐Basir OA. 2023. Deep representation of imbalanced spatio-temporal traffic flow data for traffic accident detection. IET Intelligent Transport Systems 17(3):606−19 doi: 10.1049/itr2.12287 CrossRef Google Scholar
[26]	Li L, Lin Y, Du B, Yang F, Ran B. 2022. Real-time traffic incident detection based on a hybrid deep learning model. Transportmetrica A: Transport Science 18(1):78−98 doi: 10.1080/23249935.2020.1813214 CrossRef Google Scholar
[27]	Xie T, Shang Q, Yu Y. 2022. Automated traffic incident detection: Coping with imbalanced and small datasets. IEEE Access 10:35521−40 doi: 10.1109/ACCESS.2022.3161835 CrossRef Google Scholar
[28]	Jilani U, Asif M, Rashid M, Siddique AA, Talha SMU, et al. 2022. Traffic congestion classification using GAN-based synthetic data augmentation and a novel 5-layer convolutional neural network model. Electronics 11(15):2290 doi: 10.3390/electronics11152290 CrossRef Google Scholar
[29]	Huang Y, Wei W, Yang H, Wu Q, Xu K. 2023. Intelligent algorithms for incident detection and management in smart transportation systems. Computers and Electrical Engineering 110:108839 doi: 10.1016/j.compeleceng.2023.108839 CrossRef Google Scholar
[30]	Dabboussi AH, Jammal M. 2024. Traffic data augmentation using GANs for ITS. 2024 20^th International Conference on Distributed Computing in Smart Systems and the Internet of Things (DCOSS-IoT), Abu Dhabi, United Arab Emirates, 29 April 2024−01 May 2024. USA: IEEE. pp. 66-73 doi: 10.1109/DCOSS-IoT61029.2024.00020
[31]	Benabdallah Benarmas R, Beghdad Bey K. 2024. Improving road traffic speed prediction using data augmentation: a deep generative models-based approach. Annals of Data Science 11(6):2199−216 doi: 10.1007/s40745-023-00508-x CrossRef Google Scholar
[32]	Liao C, Chen XM. 2024. A meta-learning approach to improving transferability for freeway traffic crash risk prediction. Digital Transportation and Safety 4(1):21−30 doi: 10.48130/dts-0024-0027 CrossRef Google Scholar
[33]	Avila A M, Mezić I. 2020. Data-driven analysis and forecasting of highway traffic dynamics. Nature Communications 11(1):2090 doi: 10.1038/s41467-020-15582-5 CrossRef Google Scholar
[34]	Qu Q, Shen Y, Yang M, Zhang R. 2024. Towards efficient traffic crash detection based on macro and micro data fusion on expressways: a digital twin framework. IET Intelligent Transport Systems 18(12):2725−43 doi: 10.1049/itr2.12498 CrossRef Google Scholar
[35]	Yang K, Quddus M, Antoniou C. 2022. Developing a new real-time traffic safety management framework for urban expressways utilizing reinforcement learning tree. Accident Analysis & Prevention 178:106848 doi: 10.1016/j.aap.2022.106848 CrossRef Google Scholar
[36]	Zaitouny A, Fragkou AD, Stemler T, Walker DM, Sun Y, et al. 2022. Multiple sensors data integration for traffic incident detection using the quadrant scan. Sensors 22(8):2933 doi: 10.3390/s22082933 CrossRef Google Scholar
[37]	German Aerospace Center. n.d. SUMO at a glance. SUMO Documentation. https://sumo.dlr.de/docs/SUMO_at_a_Glance.html (Retrieved December 21, 2024)
[38]	Krajzewicz D, Brockfeld E, Mikat J, Ringel J, Rössel C, et al. 2005. Simulation of modern traffic lights control systems using the open source traffic simulation SUMO. Proceedings of the 3^rd Industrial Simulation Conference 2005. Berlin, Germany: EUROSIS-ETI. pp. 299–302 https://elib.dlr.de/21012
[39]	Koh SS, Zhou B, Yang P, Yang Z, Fang H, et al. 2018. Reinforcement learning for vehicle route optimization in SUMO. 2018 IEEE 20^th International Conference on High Performance Computing and Communications; IEEE 16^th International Conference on Smart City; IEEE 4^th International Conference on Data Science and Systems (HPCC/SmartCity/DSS), Exeter, UK, 28–30 June 2018. USA: IEEE. pp. 1468–73 doi: 10.1109/HPCC/SmartCity/DSS.2018.00242
[40]	Kastner KH, Pau P. 2015. Experiences with SUMO in a real-life traffic monitoring system. SUMO 2015–Intermodal Simulation for Intermodal Transport 28, Berlin. pp. 1–10 www.researchgate.net/publication/291339917
[41]	Fernandes P, Nunes U. 2010. Platooning of autonomous vehicles with intervehicle communications in SUMO traffic simulator. 13^th International IEEE Conference on Intelligent Transportation Systems, Funchal, Portugal, 19–22 September 2010. USA: IEEE. pp. 1313–18 doi: 10.1109/ITSC.2010.5625277
[42]	Shamsashtiany R, Ameri M. 2018. Road accidents prediction with multilayer perceptron (MLP) modelling case study: roads of Qazvin, Zanjan and Hamadan. Journal of Civil Engineering and Materials Application 2(4):181−92 doi: 10.22034/jcema.2018.91998 CrossRef Google Scholar

{{lists.name}}

SUMO-enhanced traffic anomaly detection: enhancing spatiotemporal transferability through simulation-based data augmentation

Abstract

Rights and permissions

References

About this article

Cite this article

Article Metrics

Access History

Other Articles By Authors