Table 3 A comprehensive comparative study of the previous works

[134]

A proposed technique to estimate the traffic flow in more populated locations

N/A

DAN, DBN, RF, and LSTM

LSTM has an accuracy of 95.2%

[135]

Anticipate trip duration between two points on a route using neural networks

Dataset obtained using Waze Live Map APIs

K-Means

Other factors such as weather conditions need to be considered to boost the efficiency and reliability of the proposed technique

[136]

A proposed strategy for short-term traffic flow forecasting

Traffic flow data generated by sensors located on road networks in Shenzhen, China

RNN and MDN

The accuracy of the proposed network is over 92%

[137]

The authors proposed a DL approach for precise traffic prediction under bad weather conditions

Two types of data sets: first, traffic data from a highway control center, and second, weather data from local monitoring stations

Traditional DBN with the topmost layer is an SVR

The proposed DBN controlled the all-time traffic prediction error by 9% and the peak time prediction error by 15%

The computation time of the upgraded DBN needs optimization

[138]

The aim was to minimize the number of vehicles stopped at all signal intersections on the road network by proposing an urban traffic light control system

Real-world traffic data provided by the Aliyun Tianchi platform

Cell transmission model (CTM)

The comparison results showed that both the proposed system and the signal control optimization technique work well

[139]

A proposed technique for creating a traffic congestion index by extracting free-stream speed and flow

Caltrans PeMS dataset

(http://pems.dot.ca.gov)

SG-CNN

The proposed technique has an MAE of 0.01578, an RMSE of 0.02264, and a MAPE of 0.06148

[140]

The authors proposed a real-time data-driven queue length prediction technique

Data from the adaptive traffic control system InSync

LSTM, SMBO

The LSTM model was compared with other models, such as SVR, KNN, RNN, and FFNN. The RMSE and

MAE values for the LSTM model are lower than the SVR and KNN

[141]

The authors proposed an AST-MTL model for predicting multi-horizon traffic flow and velocity at the road network scale

Sets of GPS data called OBU data, which record the distance that a lorry travels within Belgium

FNN, GCN, and GRU

This study struggles with finding the right strategy for explicitly maximizing task learning

[142]

The authors proposed FI-RNNs to extract the potential correlation between traffic context and state

A real-world traffic data set is from the Digital Roadway Interactive Visualization and Evaluation Network (DRIVE Net)

Stacked RNN and Sparse Automatic Encoder

Investigating more feature extraction and merging techniques to improve performance. Also, the examination of other influencing elements is needed

[143]

A traffic situational awareness array technology was developed to extract the spatial patterns encoded in the traffic flow

Caltrans PeMS dataset

SVR and LSTM

main observation in this study is the need to improve the network structure and parameter options

[144]

A traffic congestion model was proposed to predict the traffic of neighborhoods within an area

An abbreviated version of the San Francisco Bay Area Highway Network

LSTM and Graph-CNN

The final tests showed that the model is scalable

[145]

A traffic flow prediction technique was presented to tackle the error amplification phenomenon of classical summation methods and to improve prediction performance

Real-world traffic data was obtained by microwave sensors placed on highways in Beijing, China

The IBCM-DL model is built upon the BCM framework proposed by Wang [146]

Extra factors, such as weather conditions, traffic accidents, speed, and occupancy should be considered to improve the model's reliability

[147]

Evaluate the predictive skills of some ML models

The original data was created by Aimsun (2018)

LSTM, GRU, SRCN, and HGC-LSTM

LSTM and GRU presented fewer errors than SRCN and HGC-LSTM

[148]

Traffic speed prediction models for upstream under work area conditions were proposed

Realistic road traffic accident data from Gauteng Province, South Africa (SA)

Deep ANN and CNN

The generated CNN model should be improved in many aspects such as discovering additional sources to update the traffic volume and anticipate traffic congestion in the opposite direction of traffic

[149]

The authors proposed a hybrid neural network architecture to retrieve Spatiotemporal information from the input image and predict network congestion

Datasets collected from the Caltrans PeMS

CNN, LSTM, and Transpose-CNN

Additional factors such as weather information (rain, snow, and fog) must be addressed. Also, more information from many data sources must be added to get more accurate forecasts

[150]

The authors suggested a traffic jam prediction technique based on correcting missing temporal and spatial information

The traffic speed data from the ITS are collected by traffic information collectors installed on the roads or along the roadsides

LSTM

the proposed method was found to achieve better performance with a difference in the MAPE of 3%–17%

The forecasting precision should be enhanced

Also, the authors need to build a model with improved user performance

[151]

The authors suggested the DELA technique that could help explicitly learn accurate traffic information, road structure, and weather conditions

Traffic flow information for approximately 3 months (from July 19, 2016, to October 17, 2016) provided by KDD CUP 2017

CNN and LSTM

Poor explanatory power for the selected DL models

A limited learning ability of the embedded component

[119]

A technique for large-scale, faster, and real-time traffic forecasting

11 years of data provided by the California Department of Transportation (Caltrans) [152]

Big Data, In-memory computing, GPUs, and DL

The suggested approach has poor prediction accuracy

[153]

A distinctive traffic prediction approach with the least prediction error

Real-world traffic big data of PeMS

LSTM

The model training time needs to be regulated

The number of optimized parameters needs to be expanded

[154]

A pathway-based DL framework that can provide superior traffic velocity forecasts on a citywide scale was presented

Data from Automated Vehicle Identification (AVI) detectors in the core area of Xuancheng, China

Bi-LSTM

The model was reasonable and interpretable in the urban transportation context raising the interpretability of a DL model for a transport application is an issue

[155]

Forecasting traffic congestion level using refined GPS trajectory data and the Hidden Markov model

GPS trajectory data

CNN, RNN, LSTM, GRU, ARIMA, SVR, and Ridge Regression

Insufficient GPS trajectory data was collected. More GPS data must be taken into account. The network structure can be altered to improve model performance

[156]

A spatiotemporal model for the short-term prediction of the crowding level

Traffic congestion data of Helsinki, Finland collected using HERE Traffic API

ConvLSTM

External parameters, such as points of interest, weather, and the surrounding environment should be incorporated

[157]

A DL-based methodology for directly forecasting traffic status based on a time–space diagram

Simulated data and a real-world dataset (NGSIM US-101)

CNN

Compared with SVR, MLP, and ARIMA, the suggested CNN model provided a higher generalization in traffic state prediction

[158]

Predict traffic flow more accurately based on fuzzy CNN

The Beijing taxi real route and meteorological data

F-CNN

Additional influential aspects in traffic flow forecasting need to be explored and more efficient DL models need to be used

[159]

A model for short-term traffic forecasting was proposed

Traffic flow data of the Xiangjiang Middle Road in Changsha City except the weekend of May 20th, 2018, for a total of 40 days

Spatiotemporal analysis and the GRU

Other factors such as weather conditions need to be considered