2020-01-0729 Published 14 Apr 2020

Vehicle Velocity Prediction Using Artificial Neural Network and Effect of Real World Signals on Prediction WindowTushar Gaikwad Western Michigan University

Aaron Rabinowitz Colorado State University

Farhang Motallebiaraghi Western Michigan University

Thomas Bradley Colorado State University

Zachary Asher, Alvis Fong, and Rick Meyer Western Michigan University

Citation: Gaikwad, T., Rabinowitz, A., Motallebiaraghi, F., Bradley, T. et al., “Vehicle Velocity Prediction Using Artificial Neural Network and Effect of Real World Signals on Prediction Window,” SAE Technical Paper 2020-01-0729, 2020, doi:10.4271/2020-01-0729.

Abstract

Prediction of vehicle velocity is important since it can realize improvements in the fuel economy/energy efficiency, drivability, and safety. Velocity prediction

has been addressed in many publications. Several references considered deterministic and stochastic approaches such as Markov chain, autoregressive models, and artificial neural networks. There are numerous new sensor and signal tech-nologies like vehicle-to-vehicle and vehicle-to-infrastructure communication that can be used to obtain inclusive datasets. Using these inclusive datasets of sensors in deep neural networks, high accuracy velocity predictions can be achieved. This research builds upon previous findings that Long Short-Term Memory (LSTM) deep neural networks provide low error velocity prediction. We devel-oped an LSTM deep neural network that uses different groups of datasets collected in Fort Collins, Colorado. Synchronous data was gathered using a test vehicle equipped with sensors to measure ego vehicle position and velocity,

ADAS-derived near-neighbor relative position and velocity, and infrastructure-derived transit time and signal phase and timing. The effect of different groups of input datasets on forward velocity prediction windows of 10, 15, 20, and 30 seconds was studied. The developed algorithm was tested on an NVIDIA DRIVE PX2. This research shows that the lowest Mean Absolute Error (MAE) of future velocity prediction is with a fully inclusive dataset in 10-second velocity prediction windows. It was observed that GPS data, current vehicle velocity data, and vehicle-to-infrastructure data were the most influential parameters for prediction accuracy. Additionally, we  have demonstrated that the LSTM neural network used for velocity prediction can be implemented in real-time using an NVIDIA DRIVE PX2. Integration of velocity prediction into fuel economy strate-gies and autonomous vehicle technology have the potential to improve fuel economy and safety. Future work involves demonstrating these two use cases in a physical vehicle using NVIDIA DRIVE PX2.

1.  Introduction

The shift that we  are witnessing toward Intelligent Transportation Systems (ITS) will be the most disrup-tive since the initial days of automobiles [1]. It has the

potential to completely transform the movement of people and goods, enabling safer and smarter transportation [2]. The ITS consists of several technologies such as Advanced Driver Assistance System (ADAS), Automated Driving Functions (ADF), Vehicle to Vehicle (V2V) and Vehicle to Infrastructure (V2I) communication [3]. With ITS, we can obtain informa-tion such as ego vehicle position and velocity, ADAS-derived near-neighbor relative position and velocity, and infrastruc-ture-level transit time and signal phase and timing (SPaT).

Their growing use will yield significant improvements in fuel economy (FE), transportation efficiency, and safety [4].

One of the important and emerging applications for velocity prediction is for developing an energy management strategy for fuel economy in Hybrid Electric Vehicles (HEVs) and Plug-In Hybrid Electric Vehicles (PHEVs). For increasing the economy of vehicle operation, automation, and future prediction horizon can be used with HEVs and PHEVs. Using automation and future prediction horizon with HEVs and PHEV’s, the operations of a vehicle can be  controlled. Prediction of future driving conditions essentially determines the constraints of the energy optimization problem. Methods such as detection algorithm based on the speed profile are

Downloaded from SAE International by Zachary Asher, Wednesday, April 15, 2020

VEHICLE VELOCITY PREDICTION USING ARTIFICIAL NEURAL NETWORK AND EFFECT OF REAL WORLD SIGNALS 2

also being developed [5, 6]. An inaccurate prediction can dete-riorate energy savings or even cause safety concerns. Due to such dependency between prediction horizon and Optimal EMS and detection algorithms, there is a pressing need to develop accurate and robust approaches for predicting vehicle speed and understanding which input signals would produce a better prediction. Artificial Neural Networks (ANN) have the capability to model and extract unseen features and rela-tionships, which makes it reliable to model complex prediction problems and patterns. Its ability to predict future output of systems enables the use of optimal control principles to give an optimal solution [7]. Self-driving vehicle technology poten-tial has been recognized by many researchers for informing an ANN prediction model whose outputs can derive a control strategy to improve FE [8, 9, 10, 11, 12, 13, 14, 15, 16, 17]. An important point to note is that each study uses “Intelligent Vehicle” technology inputs to a prediction model to inform optimal control.

Vehicle Speed Prediction has been addressed in many different publications. Prediction using ego vehicle velocity over 1-10 sec were compared with respect to parametric and non-parametric models in Lefèvre et al. [18]. Results show that simple models performed well for short term prediction and advanced models for long term prediction. Deep Neural Networks (DNNs) is used to predict in Olabiyi et al. [19]. Deep learning networks are also used to predict ego vehicle velocity and route in Lemieux et al. [20]. Zhang et al. utilized V2V and V2I communications for future vehicle velocity prediction [21]. They also developed an energy management strategy based on vehicle velocity prediction. Radial Basis Function neural networks predict vehicle speed performance on four standard driving cycles in Sun et  al. [22]. Hellström and Jankovic proposed a model for a human driver operating an accelerator pedal and used it for prediction [23].

Liu Kuan [24] focused on comparing different approaches predicting 10-sec forward prediction. Among all approaches, the Long Short-Term Memory (LSTM) deep neural network has been shown to the best performance. It is a powerful type of neural network, used in deep learning because very large architectures can be successfully trained. Despite the fact that the issue of vehicle speed prediction has been previously studied, the dependency of prediction on different groups of signals still needs to be understood.

Considering the potential in the accuracy of applications and implementations of vehicle velocity predictions, we used LSTM for predicting the ego vehicle velocity. We gathered different signals along the roads of Fort Collins, Colorado, and created different input groups from those signals to understand the effect on the prediction horizon. We also tested the deployment of a prediction algorithm on fast processing hardware, which would pave the way for the imple-mentation of this paper in an actual vehicle. NVIDIA Drive PX2 is such hardware, which is used to develop an algorithm for autonomous vehicles. The article is organized as follows: In section 2, the problem statement is described with a descrip-tion of the collected dataset, which is subsequently used in LSTM. It also elaborates on LSTM deep NNs. We also intro-duced NVIDIA Drive PX2 and its functioning in this section. Section 3 presents the assessment methods and the results in different scenarios. Finally, section 4 summarizes the

conclusions and sets the direction of future research in terms of implementation using NVIDIA Drive PX2.

2.  MethodologyThe objectives of this paper are to assess the potential for radar signal, ego vehicle sensor data, ADAS-derived near-neighbor relative position, infrastructure-level transit time (Segment Speed), and signal phase and timing (SPaT) to improve vehicle velocity prediction. We are also focusing on the effect of these signals on different forward prediction windows. In order to obtain features for vehicle velocity prediction, drive cycles are collected by driving a vehicle equipped with different sensors in Fort Collins, Colorado.

2.1.  Drive Cycle Development and Signal Recording

The Fort Collins dataset was collected in August 2019 and contained data from repeated drives along a fixed route by the same driver. This route represents the round trip on the following:

1. Parking Lot 2. West on Mulberry until Shields 3. South on Shields until Prospect 4. East on Prospect until College 5. North on College until Mulberry 6. West on Mulberry until Parking Lot 7. Parking Lot

This route is shown in the birds-eye view in Figure 1, with the black arrows representing the driving directions, the black circle representing the start/endpoint, the red targets repre-senting traffic signals, and the blue lines representing traffic segments. In total, this route is 4 miles and should take 10-12 minutes per cycle, depending on traffic.

 FIGURE 1  Drive cycle map of the Fort Collins dataset©

SA

E In

tern

atio

nal.

Downloaded from SAE International by Zachary Asher, Wednesday, April 15, 2020

3VEHICLE VELOCITY PREDICTION USING ARTIFICIAL NEURAL NETWORK AND EFFECT OF REAL WORLD SIGNALS

The test will generate the following data: Autonomous driver assistance system (ADAS) data for the vehicle forward cone from smart radar, which is a typical setup for production vehicle ADAS systems, and vehicle to infrastructure (V2I) data in the form of traffic signal information and segment travel times. The ego vehicle was specially instrumented research vehicle with the afore-mentioned ADAS sensors and a stereo camera with a Freematics logger. The following are some of the feature plots which are shown.

Figure 2 shows the data plot for velocity vs. time for one drive instance. The plot of longitude versus latitude versus time is shown in Figure 3. The velocity is on the map against time for three drive instances. We can observe that the velocity profiles are a similarity.

The radar objects detections immediately in front are plotted and shown in Figure 4, which shows the objects detected and position from the vehicle. Figure 5. shows the plot for the Longitude vs. Latitude vs. Altitude. Velocity is shown on this map terrain.

2.2.  Prediction Model Derivation: Long Short-Term Memory (LSTM) Deep Neural Network

Introduction to LSTM Model LSTM is a special type of Recurrent Neural Network (RNN) capable of learning long-term dependencies from large data sets (i.e., data sets with lots of inputs) and avoiding issues such as gradient vanishing/exploding of conventional RNNs. For the analysis of sequen-tial data, LSTM is the most commonly used deep learning model. An RNN composed of LSTM units is often called the LSTM network, where units of RNN are LSTM units.

Inside the repeating parts, the conventional RNNs have only one activation function, while LSTMs have three gates with different activation functions, interacting with each other. A typical LSTM unit, often called memory cell, memory block, or cell, is composed of a cell, an input gate, an output

 FIGURE 2  Velocity vs. time for one drive instance.

© S

AE

Inte

rnat

iona

l.

 FIGURE 3  Velocity on the longitude vs. latitude vs. time

© S

AE

Inte

rnat

iona

l.

 FIGURE 4  Radar object detections immediately in front.

© S

AE

Inte

rnat

iona

l.

 FIGURE 5  Velocity on the map terrain.©

SA

E In

tern

atio

nal.

Downloaded from SAE International by Zachary Asher, Wednesday, April 15, 2020

VEHICLE VELOCITY PREDICTION USING ARTIFICIAL NEURAL NETWORK AND EFFECT OF REAL WORLD SIGNALS 4

gate, and a forget gate. The cell remembers values over arbi-trary time intervals, and the three gates regulate the flow of information into and out of the cell.

The typical LSTM Cell internal structure is shown in Figure 6, which contains three gates - forget gate, input gate, and output gate. Computation details can be observed in the literature [15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27]. The gates are computation can be  summarized by the following equations.

1. Forget Gate: Decides what information to discard from the cell.

f W h X bt f tT

tT T

f= éë ùû +( )-s 1, (1)

2. Input Gate: Decides which values from the input to update the memory state.

i W h X bt i tT

tT T

i= éë ùû +( )-s 1, (2)

�C W h X bt C tT

tT T

C= éë ùû +( )-tanh 1, (3)

3. Output Gate: Decides what to output based on input and the memory of the cell.

o W h X bt o tT

tT T

o= éë ùû +( )-s 1, (4)

C f C i Ct t t t t= ´ + *-1� (5)

h o Ct t t= ´ ( )tanh (6)

where Wf, Wi, WC, Wo and bf, bi, bC, bo designate vectors (matrices) of weights and biases for forget gate, input gate, cell state, and output gate, respectively. In the above expressions, σ is a sigmoid function.

LSTM Applied to Fort Collins Data Set The develop-ment of LSTM is based on Keras 1.0 in Python 3.7. By trial and error, the structure of LSTM is chosen as a 2-layer neural

network with 3 hidden layers. The 3 hidden layers have {50,30,15} neurons. The activation function is chosen as “ReLu”, not only to avoid gradient vanishing but also to accel-erate the training process. The regularization and dropout layer are considered to avoid overfitting.

The 10 sec ahead prediction results are shown in Figure 7. The resulting LSTM model shows very promising prediction results. The effect of different signals on the prediction window is further explained in the Result section.

Assessment For accurate assessment of the prediction results, we introduced two assessment criteria: Mean absolute error (MAE) and Timeshift.

MAE: Mean Absolute Error (MAE) is a measure of the difference between two continuous variables. Assume yt1, …, ytn are prediction results and Zt1, …, Ztn are target values at time instance t. The MAE is given by,

MAE Y Zy Z

nt t

i

n

ti ti

,( ) =-

=å 1

Smaller MAE means smaller errors between prediction results and target values.

Time Shift The time shift is a measure of the time lag between the predicted time series and target time series. It uses a cross-correlation technique for finding the time shift error. A smaller time shift means a smaller time lag between the prediction results and target values. It can be expressed as,

time shift Y Zn

t t | |= *æ

èçç

ö

ø÷÷å -arg maxd d

0

1

2.3.  Implementation of Prediction Model Using In-Vehicle Hardware

The driving mode-specific performance operated by an auto-mated driving system of all aspects of the dynamic driving task. For using artificial intelligence in autonomous vehicles,

 FIGURE 6  LSTM Cell internal structure.

© S

AE

Inte

rnat

iona

l.

 FIGURE 7  LSTM model 10 sec ahead prediction results on Fort Collins dataset (Group D)

© S

AE

Inte

rnat

iona

l.

Downloaded from SAE International by Zachary Asher, Wednesday, April 15, 2020

5VEHICLE VELOCITY PREDICTION USING ARTIFICIAL NEURAL NETWORK AND EFFECT OF REAL WORLD SIGNALS

we need a supercomputer that could take level 3 level 4 opera-tion. NVIDIA Drive PX2 has the capability of taking the task of up to level 5 operation. Figure 8 Show hardware of NVIDIA Drive Px2. It can be  used for training a neural network. We can also write algorithms/applications on pc or system and port it into an embedded platform. Using all these together, we can put it into a simulated environment and drive a simulated vehicle [28]. NVIDIA DRIVE PX2 has a typical usage of 80W with the operation of 8-12 cameras radar, 2 Tegra processors, and 3GPUS. Which makes it perfect hardware to implement these strategies.

3.  Results

3.1.  Dataset Groups and Prediction Results

For comparing different signals, we divided the signals into different groups. The prediction plots show a comparison of prediction against the Target values, i.e., next prediction horizon (e.g., 10, 15, 20 30 Seconds).

Figure 9 shows the results for the velocity prediction for the 10th second using Group A signals and plotted against

the target value. It can be observed that the prediction results are not aligning with the target.

Figure 10 shows that Group B signals perform better than Group A due to the adding of the previous 5 seconds and EGO and Engine parameters Dataset.

Figure 11 shows the plots of prediction using the Group C dataset.

Figure 11 shows the prediction results using the Group D dataset. It shows significant improvement over the Group B and Group C dataset due to the usage of Spat data along with the Group B dataset.

Figure 12 shows the prediction results using the Group E dataset. It shows some mispredictions over the Group D dataset.

 FIGURE 8  Conditional Automation

© S

AE

Inte

rnat

iona

l.

 FIGURE 9  10th-second Prediction Results using Group A signals

© S

AE

Inte

rnat

iona

l.

 FIGURE 10  10th-second Prediction Results using Group B signals

© S

AE

Inte

rnat

iona

l.

 FIGURE 11  10th-second Prediction Results using Group C signals

© S

AE

Inte

rnat

iona

l.

Downloaded from SAE International by Zachary Asher, Wednesday, April 15, 2020

VEHICLE VELOCITY PREDICTION USING ARTIFICIAL NEURAL NETWORK AND EFFECT OF REAL WORLD SIGNALS 6

Figure 13 shows the velocity prediction for the 10th second using Group F signals and plotted against the target value.

Figure 14 shows the velocity prediction for the 10th second using Group G signals and plotted against the target value.

3.2.  Effect on Prediction Horizon

Signal Groups are used to obtain the prediction results for the next 10, 15, 20, and 30 seconds of forward prediction. These

results are assessed based on MAE, and the plot is shown in Figure 15. Table 1 represents the average result values for MAE. We can observe that as the target horizon increases, the MAE starts increasing. Group D results perform the best, as SPat data is beneficial for prediction. It was observed that Group B signals perform better than Group A due to addition of the previous 5 seconds of vehicle velocity and EGO and Engine parameters.

 FIGURE 13  10th-second Prediction Results using Group F signals

© S

AE

Inte

rnat

iona

l.

 FIGURE 14  10th-second Prediction Results using Group G signals

© S

AE

Inte

rnat

iona

l.

 FIGURE 11  10th-second Prediction Results using Group D signals

© S

AE

Inte

rnat

iona

l.

 FIGURE 12  10th-second Prediction Results using Group E signals

© S

AE

Inte

rnat

iona

l.

Downloaded from SAE International by Zachary Asher, Wednesday, April 15, 2020

7VEHICLE VELOCITY PREDICTION USING ARTIFICIAL NEURAL NETWORK AND EFFECT OF REAL WORLD SIGNALS

Group D data contains current velocity, GPS, previous 5 seconds of velocity, EGO, and Engine parameters, and Spat (Signal Phase and Timing). In the case of Group E and Group F, Segment speed tends to worsen prediction. Hence group E and group F show some mispredictions when compared with the group D performance. These are consistent with the results obtained in Asher, Zachary D [29], where less fuel economy was observed due to mispredictions when the Segment Speed signal is used for prediction. Overall, we observed that Group D shows an almost 50% error reduc-tion compared to Group A. Group D as an input to LSTM used for 10th-second prediction MAE of 1.78 m/s and the time shift of 2.25 seconds is observed.

Table 1 shows the Performance of Group Signals on the prediction horizon.

The effect due to signal groups on time shift is shown in Figure 16. The average result values are shown in Table 2. We can observe that, as the prediction window increases, the time shift in the prediction increases.

3.3.  Velocity Prediction In-Vehicle Implementation

For the simulation of the results, we used a desktop computer installed with Python 3.7, Keras - TensorFlow, with RTX 2060 GPU. We first collected data and then used to derive the results. NVIDIA PX2 can gather the live data while mounted in the vehicle, which can perform multiple highly tasked computations. We implemented our algorithm in a similar way on the NVIDIA PX2 architecture, which is a different

environment, where we observed the same results. Overall, we can observe the possibility of having predicted the same results from different architecture and different environments, which would pave the path for the future implementation of this algorithm in the actual vehicle.

4.  ConclusionsIn this study, multiple signals collected in Fort Collins, Colorado, which are used in different Groups as an input feature for LSTM. LSTM shows good performance for predicting vehicle velocity and can be used for predicting different prediction horizons. Assessment of results was done with MAE and compared the Predictions from different Groups. An increase in the error was observed with an increase in the prediction horizon. Group D dataset showed the best performance, where SPaT data was observed very useful for prediction. Although Group E has more signals, the Segment speed did not improve the results. Group D as an input to LSTM used for 10th-second prediction MAE of 1.78 m/s and the time shift of 2.25 seconds is observed. We can also observe that Group D shows almost 50% error reduction compared to Group A. Finally, we implemented the algorithm on the vehicle hardware NVIDIA Drive PX2, where we observed similar results.

This study demonstrates that V2I data is very important to obtain significant improvement in vehicle velocity predic-tion. ANN’s ability to predict future velocity can be observed.

 FIGURE 15  Effect of signal groups on prediction window MAE

© S

AE

Inte

rnat

iona

l.

TABLE 1 Effect of signal groups on prediction window MAE

Signal GroupsPrediction Window10 sec 15 sec 20 sec 30 sec

Group A 2.45 3.52 4.32 5.29

Group B 2.14 3.22 4.02 4.91

Group C 2.18 3.24 4.02 5.02

Group D 1.78 2.55 3.09 4.04

Group E 1.86 2.64 3.24 4.18

Group F 1.85 2.66 3.28 4.23

Group G 2.14 3.11 3.71 4.53© S

AE

Inte

rnat

iona

l.

 FIGURE 16  Effect of signal groups on prediction window Time shift

© S

AE

Inte

rnat

iona

l.

TABLE 2 Effect of signal groups on prediction window time shift

Signal GroupsPrediction Window10 sec 15 sec 20 sec 30 sec

Group A 5.67 9.65 13.43 22.19

Group B 4.23 7.59 10.64 17.31

Group C 4.25 7.55 10.72 17.48

Group D 2.25 3.68 4.05 7.07

Group E 2.42 3.23 4.27 7.45

Group F 2.35 3.44 4.23 7.46

Group G 5.67 9.65 13.43 22.19© S

AE

Inte

rnat

iona

l.

Downloaded from SAE International by Zachary Asher, Wednesday, April 15, 2020

VEHICLE VELOCITY PREDICTION USING ARTIFICIAL NEURAL NETWORK AND EFFECT OF REAL WORLD SIGNALS 8

It also shows the performance of different signals on the prediction using the LSTM neural network. In future work, the prediction method could be extended with different NN architectures with a greater number of drive cycles. It also seeks work on implementation where live signals collected on NVIDIA Drive PX2 can be used for getting live predictions.

References 1. Fagnant, D.J. and Kockelman, K., “Preparing a Nation for

Autonomous Vehicles: Opportunities, Barriers and Policy Recommendations,” Transportation Research Part A: Policy and Practice 77:167-181, 2015.

2. Waschl, H., Kolmanovsky, I., and Willems, F., Control Strategies for Advanced Driver Assistance Systems and Autonomous Driving Functions (Springer Verlag, 2019).

3. Birky, A., Laughlin, M., Tartaglia, K., Price, R. et al., “Electrification Beyond Light Duty: Class 2b-3 Commercial Vehicles,” No. ORNL/TM-2017/744, Oak Ridge National Lab. (ORNL), Oak Ridge, TN, 2018.

4. Nkoro, A.B. and Vershinin, Y.A., “Current and Future Trends in Applications of Intelligent Transport Systems on Cars and Infrastructure,” in IEEE 17th International Conference on Intelligent Transportation Systems, 2014, doi:10.1109/ITSC.2014.6957741.

5. Dadras, S., Jamshidi, H., Dadras, S., and Pilutti, T.E., “Novel Stop Sign Detection Algorithm Based on Vehicle Speed Profile,” in 2019 American Control Conference (ACC), 2019, 3994-3999, IEEE.

6. Dadras, S., “Path Tracking Using Fractional Order Extremum Seeking Controller for Autonomous Ground Vehicle,” SAE Technical Paper 2017-01-0094, 2017. https://doi.org/10.4271/2017-01-0094.

7. Demuth, H.B., Beale, M.H., De Jess, O., and Hagan, M.T., Neural Network Design Second Edition (Martin Hagan: USA, 2014).

8. Bender, F.A., Kaszynski, M., and Sawodny, O., “Drive Cycle Prediction and Energy Management Optimization for Hybrid Hydraulic Vehicles,” IEEE Trans Veh Technol 62:3581-3592, 2013.

9. Donateo, T., Pacella, D., and Laforgia, D., “Development of an Energy Management Strategy for Plug-In Series Hybrid Electric Vehicle Based on the Prediction of the Future Driving Cycles by ICT Technologies and Optimized Maps,” SAE Technical Paper 2011-01-0892, 2011. https://doi.org/10.4271/2011-01-0892.

10. Sun, C., Moura, S.J., Hu, X., Hedrick, J.K., and Sun, F., “Dynamic Traffic Feedback Data Enabled Energy Management in Plug-In Hybrid Electric Vehicles,” IEEE Trans Control Syst Technol 23:1075-1086, 2015.

11. Sun, C., Hu, X., Moura, S.J., and Sun, F., “Velocity Predictors for Predictive Energy Management in Hybrid Electric Vehicles,” IEEE Trans Control Syst Technol 23:1197-1204, 2015.

12. Sun, C., Hu, X., Moura, S.J., and Sun, F., “Velocity Predictors for Predictive Energy Management in Hybrid Electric Vehicles,” IEEE Trans Control Syst Technol 23:1197-1204, 2015.

13. Gong, Q., Li, Y., and Peng, Z., “Power Management of Plug-In Hybrid Electric Vehicles Using Neural Network Based trip Modeling,” in 2009 American Control Conference, 2009, 4601-4606.

14. Rezaei, A. and Burl, J.B., “Prediction of Vehicle Velocity for Model Predictive Control,” IFAC-PapersOnLine 48:257-262, 2015.

15. Gong, Q., Li, Y., and Peng, Z.R., “Trip-Based Optimal Power Management of Plug-in Hybrid Electric Vehicles,” IEEE Trans Veh Technol 57:3393-3401, 2008.

16. Mohd Zulkefli, M.A., Zheng, J., Sun, Z., and Liu, H.X., “Hybrid Powertrain Optimization with Trajectory Prediction Based on Inter-Vehicle-Communication and Vehicle-Infrastructure-Integration,” Transp Res Part C: Emerg Technol 45:41-63, 2014.

17. Gaikwad, T.D., Asher, Z.D., Liu, K., Huang, M., and Kolmanovsky, I., “Vehicle Velocity Prediction and Energy Management Strategy Part 2: Integration of Machine Learning Vehicle Velocity Prediction with Optimal Energy Management to Improve Fuel.”

18. Lefèvre, S., Sun, C., Bajcsy, R., and Laugier, C., “Comparison of Parametric and Non-Parametric Approaches for Vehicle Speed Prediction,” in American Control Conference, June 4-6, 2014, doi:10.1109/ACC.2014.6858871.

19. Olabiyi, O., Martinson, E., Chintalapudi, V., and Guo, R., “Driver Action Prediction Using Deep (Bidirectional) Recurrent Neural Network,” eprint arXiv:1706.02257, 2017.

20. Lemieux, J. and Ma, Y., “Vehicle Speed Prediction Using Deep Learning,” in IEEE Vehicle Power and Propulsion Conference, 2015, 1-5.

21. Zhang, F., Xi, J., and Langari, R., “Real-Time Energy Management Strategy Based on Velocity Forecasts Using V2V and V2I Communications,” IEEE Transactions on Intelligent Transportation Systems 18:416-430, 2017.

22. Sun, C., Sun, F., Hu, X., Hedrick, J., et al., “Integrating Traffic Velocity Data into Predictive Energy Management of Plug-In Hybrid Electric Vehicles,” in American Control Conference, Chicago, IL, 2015.

23. Hellstrom, E. and Jankovic, M., “A Driver Model for Velocity Tracking with Look-Ahead,” in American Control Conference, Chicago, IL, 2015.

24. Liu, K., Asher, Z., Gong, X., Huang, M., and Kolmanovsky, I., “Vehicle Velocity Prediction and Energy Management Strategy Part 1: Deterministic and Stochastic Vehicle Velocity Prediction Using Machine Learning,” SAE Technical Paper 2019-01-1051, 2019. https://doi.org/10.4271/2019-01-1051.

25. Olah, C. et al., “Understanding LSTM Networks,” 2015, http://colah.github.io/posts/2015-08-Understanding-LSTMs/.

26. “LSTM and RNN Tutorial with Demo (with Stock/Bitcoin Time Series Prediction, Sentiment Analysis, Music Generation),” https://github.com/omerbsezer/LSTM_RNN_Tutorials_with_Demo.

27. Huang, C.-J. and Kuo, P.-H., “A Deep CNN-LSTM Model for Particulate Matter (PM2. 5) Forecasting in Smart Cities,” Sensors 18(7):2220, 2018.

Downloaded from SAE International by Zachary Asher, Wednesday, April 15, 2020

© 2020 SAE International. All rights reserved. No part of this publication may be reproduced, stored in a retrieval system, or transmitted, in any form or by any means, electronic, mechanical, photocopying, recording, or otherwise, without the prior written permission of SAE International.

Positions and opinions advanced in this work are those of the author(s) and not necessarily those of SAE International. Responsibility for the content of the work lies solely with the author(s).

ISSN 0148-7191

9VEHICLE VELOCITY PREDICTION USING ARTIFICIAL NEURAL NETWORK AND EFFECT OF REAL WORLD SIGNALS

28. “NVIDIA DRIVE Developer Program for DRIVE PX 2,” https://developer.nvidia.com/drive/documentation.

29. Asher, Z.D., Tunnell, J.A., Baker, D.A., Fitzgerald, R.J. et al., “Enabling Prediction for Optimal Fuel Economy Vehicle Control,” SAE Technical Paper 2018-01-1015, 2018. https://doi.org/10.4271/2018-01-1015.

Contact InformationZachary D. Asher, Ph. D.Western Michigan University, Kalamazoo, MI 49008, USAzach.asher@wmich.edu

AcknowledgmentThis material is based upon work supported by the U.S. Department of Energy’s Office of Energy Efficiency and Renewable Energy (EERE). The specific organization over-seeing this report is the Vehicle Technologies Office under award number DE-EE0008468.

Definitions/AbbreviationsLSTM - Long Short-Term MemoryADAS - Advanced Driver Assistance SystemANN - Artificial Neural NetworksCAV - connected and autonomous vehiclesDNN - Deep Neural NetworksFE - fuel economyHEV - Hybrid Electric VehiclesITS - Intelligent Transportation SystemsMAE - Mean Absolute ErrorOptimal EMS - Optimal Energy Management SystemsPHEV - Plug-In Hybrid Electric VehiclesRNN - Recurrent Neural NetworkSPaT - signal phase and timingV2I - Vehicle to InfrastructureV2V - Vehicle to Vehicle

Downloaded from SAE International by Zachary Asher, Wednesday, April 15, 2020