Applications of deep learning in traffic congestion detection, prediction and alleviation: A survey

• The keywords identify deep learning, transportation, congestion, recurring and non-recurring congestion, and accidents as the survey’s topics.The passage also contains the affiliation line “Stefano-Franscini-Platz 5.”Source: §1 · p. 1

1 Introduction

• Congestion decreases road-network serviceability and is associated with lost work hours, driver aggression, accident risk, and higher greenhouse-gas emissions.The cited estimate for the USA was 8.8 billion working hours lost to congestion in one year.Source: §1 · p. 1
• Congestion prediction is described as more difficult than traffic prediction during uncongested conditions.The survey covers deep learning for detection, prediction, and alleviation, separating recurring and non-recurring congestion and identifying research gaps and future directions.Source: §1 · p. 1

2 Preliminaries

• The preliminaries target researchers in transportation and deep learning and introduce terminology used throughout the survey.Source: §2 · p. 2

2.1 Relevant concepts and terms in deep learning

• A neuron maps input features through weights and a nonlinear activation to an output; supervised training fits this mapping by minimizing loss and checking generalization on unseen data.The survey notes that transportation-informed generalization techniques receive special attention.Source: §2.1 · p. 2, §2.1.1 · p. 3

2.1.1 Fully connected layered neural networks

• A feed forward neural network (FFNN) stacks layers so each layer receives outputs from the preceding one; this survey uses FFNN rather than ANN for that architecture.The survey calls networks with more than one hidden layer deep, while noting that the literature gives no fixed depth threshold.Source: §2.1.1 · p. 2

2.1.2 CNNs and RNNs

• CNNs capture spatial dependencies, while RNNs maintain state across time steps to model temporal dependencies in traffic data.LSTMs can capture dependencies between data points far apart in time; GRUs are described as less complex and faster to train than LSTMs.Source: §2.1.2 · p. 3

2.1.3 Reinforcement Learning

• Deep reinforcement learning learns a policy to control a system toward a long-term numerical objective, and Q-learning is reported as popular in traffic prediction.The survey notes that computational load restricts the wide use of deep reinforcement learning.Source: §2.1.3 · p. 4

2.1.4 Commonly used metrics

• For congestion classification, accuracy can mislead under class imbalance, so the survey notes balanced accuracy as usual practice and discusses regression metrics including MAPE.With 60 congested observations among 1,440, predicting every point as not congested yields 95.8% accuracy; balanced accuracy averages the true-positive rates across classes.Source: §2.1.4 · pp. 4–5

2.2 Relevant concepts and terms in transportation

• Traffic state is commonly measured through density, speed, and flow, with flow, density, and speed related by q = k ∗u.Density has traditionally been difficult to measure directly across a road segment, while occupancy is often used as a proxy.Source: §2.2 · pp. 5–7
• Space-mean speed satisfies the fundamental flow relation, whereas time-mean speed does not; preprocessing determines which speed a prediction model targets.The research target can also reflect whether the goal is network throughput or user travel time.Source: §2.2 · pp. 6–7
• Floating car data increasingly supports travel-time estimates, and its data volume rises with traffic flow, unlike point sensors whose placement is a major challenge.The percentage of vehicles sharing data varies over time.Source: §2.2 · p. 7
• Figure 6 assumes a linear speed–density relation and marks maximum speed and flow as stakeholder-specific serviceability optima, with jam density indicating no serviceability.Jam density corresponds to zero speed; free-flow speed is the segment’s maximum speed under the stated speed-limit assumption.Source: Figure 6 · p. 7, §2.2 · p. 7
• Traffic hysteresis produces a loop in flow-density data, and differential acceleration and deceleration can leave disruption effects after the event ends.The survey identifies increasing traffic-flow modeling complexity as traffic approaches congestion.Source: §2.2 · pp. 7–8
• Microscopic agent-based simulators model drivers and infrastructure as agents, while model-based short-term prediction includes DynaMIT and DYNASMART.Behavioral models use expert-programmed agent decisions and calibrated parameters with physical significance.Source: §2.2 · p. 8

2.3 Definition and classification of congestion

• Congestion is defined as facility input volume exceeding output capacity, with serviceability losses varying according to the number and size of facilities.The literature operationalizes congestion using traffic-state variables or derived measures, then quantizes them into classification levels.Source: §2.3 · p. 8
• A congestion classification can use a threshold on a selected traffic variable, such as density, to label traffic as jammed or not jammed.Possible variables include speed, density, flow, or the ratio of average speed to the speed limit.Source: §2.3 · p. 8
• Recurring congestion repeats in space or time and is almost always attributed to infrastructural bottlenecks, while non-recurring congestion lacks such patterns and can follow unexpected or planned events.Examples include extreme weather, disasters, accidents, concerts, and roadworks.Source: §2.3 · pp. 8–9
• Many reviewed papers did not specify whether they addressed recurring or non-recurring congestion, so the authors assessed their results and assigned them to the respective sections.The survey discusses the two congestion types separately.Source: §2.3 · pp. 8–9

2.4 Synergies between model-based and deep learning based traffic prediction

• Data-driven prediction maps sensor observations to future traffic states without modeling user behavior, and deep learning’s uninterpretable weights make extensive validation necessary.The approach treats traffic as a measurable system state.Source: §2.4 · p. 9
• Traffic simulators support policy studies when real data are unavailable, including scenarios such as a city-wide traffic-light failure.Such studies rely on modeled driver behavior and physical interactions to provide reliable inferences.Source: §2.4 · p. 9
• Simulator-generated accident data have trained deep neural networks, microscopic simulators are incorporated in deep reinforcement learning for network control, and generative models have produced simulation agents.These examples address class imbalance in congestion and accident databases as well as simulation needs for control.Source: §2.4 · p. 9
• Physics-informed deep learning uses domain knowledge to regularize models or discover equations, while traffic simulators can encode traffic dynamics during training.The survey describes potential robustness and interpretability benefits but says transportation PIDL research remains nascent and is not covered; one cited comparison reports training over 50% faster with fixed sensor locations.Source: §2.4 · p. 9

3 Previous surveys and organisation of this survey

• CNNs model spatial interdependencies and RNNs temporal variations, and the surveyed traffic-prediction architectures commonly combine these building blocks.Source: §3 · p. 9
• Prior surveys span transportation tasks, spatio-temporal domains (Wang et al., 2020), GNNs (Wu et al., 2020), and city flows.Source: §3 · pp. 9–10
• The survey addresses deep learning for congestion detection, prediction, and alleviation across recurring and non-recurring cases, while congestion prediction itself is not divided by type.The authors explain that image-based congestion detection covers both types and say they incorporate policymakers’ challenges when possible.Source: §3 · p. 10
• The survey summarizes architecture and dataset choices and highlights key sensitivity-analysis insights where available, with most covered studies published from 2016 to 2021.Figure 8 marks a potential link between recurring and non-recurring congestion alleviation.Source: §3 · p. 10, Figure 8 · p. 10

4 Deep learning for congestion detection

• Detection models are not separated by recurring status because they detect both congestion types regardless of cause; studies instead vary by data source.Traffic-camera images capture all vehicles, and vehicle-loop-detector occupancy can provide automatic labels instead of manual annotation.Source: §4 · pp. 10–11
• CNN image models can be trained from scratch with very large image datasets, initialized from pretrained weights with small datasets, or extended with an added module.Applied to images, models estimate vehicle counts and traffic density; applied to video sequences, they estimate speed.Source: §4 · p. 11
• Chakraborty et al. (2018) report 90.5% accuracy for AlexNet and 91.2% for YOLO; Wang et al. (2018) report 78% for AlexNet and 81% for VGGNet.Wang et al. also report that AlexNet trains significantly faster because the network is smaller.Source: §4 · p. 11
• For vehicle detection across low-, mid-, and high-resolution images, Mask-RCNN accuracy is 46%, 89%, and 91%, while YOLO accuracy is 82%, 86%, and 91%.YOLO is reported as resistant to image quality and to train in almost half the time of Mask-RCNN.Source: §4 · p. 11
• Using YOLO detections with Resnet, Impedovo et al. (2019) report 99.7% accuracy for light, 97.2% for medium, and 95.9% for heavy congestion.Congestion is classified into three classes.Source: §4 · p. 11
• Kurniawan et al. (2018) report 89.5% average accuracy; Rashmi and Shantala (2020) report 92%–99% vehicle-counting accuracy but poor performance on zone-specific transport modes.The latter study concerns highly heterogeneous traffic in Karnataka and uses a YOLO model pre-trained on COCO.Source: §4 · p. 11
• Image quality and model performance differ by data source, while heterogeneous traffic streams in developing countries remain a major challenge.The survey identifies image super-resolution and video frame-rate increment as potential improvement avenues for camera-based detection.Source: §4 · p. 11

5.1 Deep learning for recurring congestion prediction

• Recurring congestion prediction covers daily variation in congestion timing and severity, using classification or traffic-variable regression across familiar network locations.Some regression studies are included when performance remains stable as traffic approaches congestion.Source: §5.1 · pp. 11–12
• With a 90-minute input horizon, CNN performance matched recurrent models; at 60 minutes, PredNet reached 84.2% mean accuracy versus Conv-LSTM’s 75.67%.PredNet was also reported to train about eight times faster than Conv-LSTM.Source: §5.1 · p. 12
• Performance varies across road types and prediction horizons, while attention weights were higher for higher-order downstream than upstream neighbours at 60 minutes.Reported QWK declined from 0.69 at 15 minutes to 0.52 at 60 minutes; suburban and urban MAPE were 4.297% and 6.087%, respectively.Source: §5.1 · pp. 12–13
• Threshold choice affects congestion classification, while weighted losses address poorer performance at higher congestion and erRCNN adapts to changing traffic states.For erRCNN, speed-prediction RMSE ranged from 5 km/h at a 10-minute horizon to 8 km/h at 50 minutes; segment importance was used to identify potential congestion sources.Source: §5.1 · pp. 13–14
• Congestion-tree prediction converts combined trees into adjacency-preserving spatial matrices for Conv-LSTM forecasting and reconstructs predicted trees for visualization.At five minutes, reported MSE was 0.27 on weekdays and 0.07 on weekends; the LSTM comparison was 0.59 and 0.32, respectively.Source: §5.1 · p. 14
• The survey finds few fixed-dataset, fixed-task comparisons, a reported maximum prediction horizon of 60 minutes, and few online-learning studies; it identifies data fusion as a post-accident research gap.It suggests fixed dataset benchmarks and further horizon analysis, and notes that online learning can be evaluated with temporal splits of historical data.Source: §5.1 · p. 14, §5.2 · p. 17

5.2 Deep learning for non-recurring congestion

• For accident-related congestion, studies predict post-accident conditions using approaches such as heat-map CNNs, multimodal models, and incident-aware DCRNN variants.Reported outcomes include 86.6% accuracy for binary congestion classification and 0.97% MAPE for three-hour post-incident speed prediction.Source: §5.2 · pp. 14, 16
• Clearance-time models use accident and traffic data, and some update predictions as information arrives; one severe-accident model update lowered RMSE from 10.8 to 7.62 minutes.Other reported clearance-time results include 40% MAPE and 8.3-minute RMSE; the survey notes that post-accident studies use varied fusion approaches with little consensus.Source: §5.2 · pp. 16–17

6 Deep learning for congestion alleviation

• Recurring congestion is addressed by reducing peak-demand severity through network controls, while non-recurring congestion efforts mainly target accident reduction and risk alerts.The survey focuses on network-level recurring-congestion control; planned events, weather, and natural disasters are described as better suited to traffic-simulator scenario studies.Source: §6 · p. 17, §6.1 · p. 18

6.1.1 Adaptive traffic signal control (TSC)

• Deep reinforcement learning is common in traffic-signal control, but optimized policies may trade throughput against fairness, especially for priority vehicles.Studies incorporate priority through reward penalties or person-based measures; the latter can disadvantage low-head-count vehicles such as ambulances.Source: §6.1 · p. 18, §6.1.1 · pp. 18, 20
• For peak-hour control on a synthetic single-intersection network, Genders and Razavi (2019) report that a feed-forward DQN reduced average total vehicle delay by around 40% versus traditionally used actuated signal control.Source: §6.1.1 · p. 20

6.1.2 Variable Speed Limit Control (VSLC)

• A VSL DQN increased average speed by 13.5% and decreased average traffic density by 25%; a separate DVSL model reduced average travel time by 5.8% versus no VSL.The latter study assumes ubiquitous connected and autonomous vehicles and uses lane-specific speed limits; the survey notes that real-world deployment depends on driver compliance.Source: §6.1.1 · p. 19, §6.1.2 · pp. 20–21
• The survey identifies small, homogeneous networks, synthetic demand, and limited comparison with existing controls as barriers to validating network-control research for deployment.It also notes that academic demand-side research is constrained by data availability and privacy concerns around personal mobility data.Source: §6.1.2 · p. 21

6.2 Deep learning for non-recurring congestion alleviation

• Accident-risk prediction uses grid-based classification or regression, with approaches addressing transfer across cities, spatial heterogeneity, and online learning.Reported examples include comparable cross-city accuracy, different urban and rural error levels, and weather-specific online-learning accuracies.Source: §6.2 · pp. 21–22

6.3 Potential link between non-recurring and recurring congestion alleviation

• Research has not established a universal reverse causal link from congestion to accidents, although recent work with many data points often finds a U-shaped relationship.That pattern associates both very high and very low traffic density with more accidents; if reproduced, it could connect accident reduction with congestion reduction toward maximum flow.Source: §6.3 · p. 22, Figure 9 · p. 23

7.1.1 Identifying the state-of-the-art model

• Reported performance is difficult to compare across studies because congestion definitions, task quantisations, data resolutions, prediction horizons, and evaluation metrics differ.Binary prediction is usually easier than multi-class prediction on the same dataset; higher-resolution inputs can change task difficulty, and most models’ performance degrades as prediction horizons increase.Source: §7.1.1 · pp. 23–24
• Deep learning models show a systematic performance drop on congestion-related tasks in developing countries, where traffic streams are more heterogeneous.The survey reports this trend across congestion detection, prediction, and alleviation.Source: §7.1.1 · p. 24

7.1.2 Lack of online learning

• Online learning updates models as data arrive, more closely matching deployment and eliminating the challenge of test labels being available to the modeller beforehand.The survey found few examples of online-learning models, possibly because relaying real-time traffic data raises security and privacy concerns.Source: §7.1.2 · p. 24

7.2 Conclusions and future research directions

• The proposed future directions address the identified challenges through data standardisation, cross-domain synergies, and combining simulation-based and deep-learning approaches.Source: §7.2 · p. 24

7.2.1 Standardised datasets and competitions

• Standardised datasets can improve model performance through more good-quality data, support clear state-of-the-art comparisons, enable transfer learning, and keep test data unseen during competitions.The survey notes that a trained network could be applied to a new city where traffic data might not be available.Source: §7.2.1 · p. 24

7.2.2 Similar prediction tasks in other domains

• Transportation researchers can draw on related prediction tasks in wireless networks and weather forecasting, including analogies between wireless routing and traffic signal control.The survey also identifies congestion-aware wireless routing as a counterpart to transportation congestion alleviation.Source: §7.2.2 · p. 25

7.2.3 Synergies between model-driven and deep learning based approaches

• Deep learning offers predictive power and faster, easier calibration than traffic simulators, while simulators provide interpretability and can generate congestion labels to address class imbalance.The survey points to deep-reinforcement-learning signal control and calls for more work combining the approaches.Source: §7.2.3 · p. 25