See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/338914748

Gaussian Process Auto Regression for vehicle center coordinates Trajectory

Prediction

Conference Paper · December 2019

DOI: 10.1109/TENCON.2019.8929719

CITATIONS READS

10 276

3 authors:

Kritika Johari Lim Qun

Singapore University of Technology and Design Singapore University of Technology and Design
10 PUBLICATIONS 19 CITATIONS 3 PUBLICATIONS 40 CITATIONS

SEE PROFILE SEE PROFILE

U-Xuan Tan
Singapore University of Technology and Design
182 PUBLICATIONS 3,188 CITATIONS

SEE PROFILE

All content following this page was uploaded by Kritika Johari on 09 October 2020.

The user has requested enhancement of the downloaded file.

 Gaussian Process Auto Regression for vehicle center coordinates
Trajectory Prediction
Qun Lim, Kritika Johari and U-Xuan Tan, Member, IEEE
Singapore University of Technology and Design
8 Somapah Road, Singapore 487372
Email: [email protected]; [email protected]; [email protected];

Abstract— With the increase in autonomous car technology,

and advance driver assistance systems (ADAS), the demand for
II. RELATED WORKS
vehicle trajectory prediction is increasing. These systems mostly Trajectory prediction of vehicles is the estimation of vehicle
use many sensors such as lidar, radar, stereo cameras which are coordinates in the future which is a subset of time series
big and expensive to detect the location of the ego vehicles with forecasting that has been heavily researched on. A time series
respect to other vehicles. Such sensors are not readily available to forecasting is a model which predict future data based on
Vulnerable Road Users (VRU) such as motorcyclist, cyclist as they observed data in the past. Currently, there are many time series
mostly only rely on their phones or a small device for navigation forecasting methods and this section will explore the related
and safety warnings. Trajectory prediction is important to VRU as works on time series forecasting such as Kalman filter, least
it is able to predict if the trajectory of other vehicles is a threat or square method, moving average, gaussian process regression and
not. Most of the current Trajectory prediction for vehicles involves Long Short-Term Memory (LSTM).
plotting a relative position of other vehicles with respect to the ego In [2][3][4][5], Kalman filter are used to predict trajectory of
vehicle and this is not possible with just a mobile device. Hence, this moving objects in 3d coordinates and although this paper is
paper proposes a method to predict trajectory of other vehicles successful in predicting the 3d motion, it is only able to predict
based on detected vehicles center coordinates. This is achieved trajectory short term of 1 step ahead instead of long term with h-
using Gaussian Process Auto-Regression based on past data of steps and hence, Kalman filter in this paper only serve as a short
vehicle center coordinates to predict future coordinates on an x-y term predictor. Moreover, Kalman filter requires a kinematics
pixel plane using only a camera sensor and You Only Look Once
model which may not be accurate due to assumptions such as
(YOLO) vehicle detection.
constant velocity or acceleration model and it fails to consider
Index Terms-Gaussian Process, Auto Regression, Trajectory
Prediction,
factors such as wind resistance which will affect the model. In
addition, Kalman filter also uses a measurement model which
I. INTRODUCTION requires the measured data of vehicle positioning, velocity and
Globally, the amount of Road Traffic Accidents (RTA) acceleration which must be measured using different sensors.
amounts to 1.25 million lives annually according to the World Least square method for trajectory prediction also has its
Health Organization (WHO) [1] and advances in technology are limitations as it depends on the order of function used. In [2] it
made to decrease RTA. Hence, autonomous car technologies and uses a second order function which is a good model for a
Advanced Driver Assistance System (ADAS) are being parabolic trajectory, but this will also mean that the algorithm
developed to support drivers to practice safe driving. ADAS must be able to determine the order of the function which will be
includes many features such as Forward Collision Warning changing depending on the road curvature.
System, adaptive cruise control, lane departure warning and Another common forecasting is called the moving average
Trajectory Prediction of other vehicles on the road. In this paper, which is used in [6][7]. In both [6] and [7], moving average is
the focus will be to predict trajectory of other vehicles with supported with other algorithms such classification in [6] and
Gaussian Process Auto-Regression. This is done by first auto regression in [7] to achieve its desired results because
detecting vehicles and finding their center coordinates based on moving average as a prediction algorithm is too simple and it is
YOLO detection on a video which requires only a camera sensor. only good in short term prediction. In [7], Auto-regressive
Next, these data of center coordinates will be collected into an moving average is used but it is still unable to predict non
array which will then be used for the Gaussian Process Auto- stationary data accurately and hence Empirical Mode
Regression for future center coordinates prediction which will be Decomposition (EMD) is used to first smoothen and stabilize the
the trajectory prediction. data. However, EMD also requires strict conditions for this
Trajectory Prediction of vehicle center coordinates has its method to work.
challenges as it is an extrapolation problem and not an Gaussian Process regression (GPR) can also predict
interpolation problem which means that the algorithm has to trajectory of vehicles and moving objects [8][9]. However, while
estimate the predicted center coordinates beyond the original [8] shows success in predicting trajectories of vehicle long term,
observed center coordinates. In addition, the further the gaussian process regression is mainly an interpolation prediction
algorithm predicts into the future, the larger the uncertainty in and hence, the trajectory prediction in [8] will only be limited to
prediction. Moreover, due to the complexity of the road the area where GPR is trained at such as a particular traffic
environment, trajectory prediction of other vehicles is difficult junction. The trained data are not easily transferable, and it will
because there is no fixed trend which can be applied to predict not serve as a good trajectory prediction for ADAS whereby
center coordinates. There have been many attempts to model vehicles are always on the move. In [9], the author uses gaussian
vehicle trajectory motion through the laws of physics as well as process with auto-regressive technique and it shows that gaussian
traffic laws, but these models are not able to cater for every process auto-regression is capable of predicting long term
situation as different country have different traffic laws and trajectories and it even produces better results than robust linear
physics-based algorithm requires many sensors for the algorithm regression model but this paper is based on a simulated trajectory
to work. Hence, this paper aims to develop an algorithm which of simple kinematics equations and has yet to propose a method
is able to predict trajectory of vehicle center coordinates using a to implement on vehicles for ADAS.
single camera sensor without the use of physics-based or rule- Lastly, with the increasing popularity of Deep Neural
based model. Network in the recent years there are many networks which is
used for trajectory prediction such as Recurring Neural Network IV. TRAJECTORY PREDICTION ALGORITHM WITH GAUSSIAN
(RNN) and Long Short-Term Memory (LSTM) PROCESS
[10][11][12][13]. Based on these papers, RNN and LSTM shows Gaussian Process is chosen for trajectory prediction as
decent results as compared to other state of the art but LSTM unlike Kalman filter, moving average and LSTM, it is able to
such as [13] requires plenty of sensor data to detect the ego predict future (x, y) pixel coordinates of vehicles without a
vehicle’s speed, yaw rate, other vehicles’ relative velocity and physics-based or rule-based model and it can be done with just
coordinates which is not achievable on a single camera sensor a singular camera sensor. As mentioned previously, Gaussian
platform. Process is good for interpolation in [8] but it can be used in
In this paper, we propose a method to perform trajectory
extrapolation for trajectory prediction as well. In this section, we
prediction of vehicles based on a singular camera platform using
will explain Gaussian Process in detail and explore long term
Gaussian Process Auto Regression and we will compare it with
existing methods. The rest of the paper will be organized as trajectory prediction with Gaussian Process Regression and
follows. In Section III, we will briefly explain proposed Gaussian Process Auto Regression.
algorithm flow. In Section IV, we will describe the proposed A. Gaussian Process Regression (GPR)
trajectory prediction method in detail. In Section V, we will show
the experimental results with comparison to other methods and Gaussian Process (GP) is defined as a collection of random
Section VI concludes the paper. variables, any finite number of which have a joint Gaussian
distribution [15]. This means that the distribution of Gaussian
Process is the joint distribution of these random variables and
III. THE PROPOSED ALGORITHM hence, it is a distribution over functions with a continuous
domain. Mathematically, a Gaussian Process is defined by its
mean function 𝑚(𝑥) and covariance function 𝑘(𝑥, 𝑥 ′ ):
𝑓(𝑥)~𝐺𝑃(𝑚(𝑥), 𝑘(𝑥, 𝑥 ′ ) (1)
By introducing the Bayes’ rule, we can train the model to get
the posterior predictive distribution to predict new values of
output y. According to Bayes’ rule, the posterior probability is
proportional to the multiplication of the likelihood of observing
y given a function f and the prior probability as follows:
𝑝(𝒇|𝑿, 𝒚) ∝ 𝑝(𝒚|𝒇)𝑝(𝒇|𝑿) (2)
In equation (2), there are 2 terms that needs to be definite
and they are the likelihood function, 𝑝(𝒚|𝒇), and prior over
Figure 1: Flow chart of GPAR Trajectory Prediction functions, 𝑝(𝒇|𝑿). The likelihood function can be defined as
follows:
In order to obtain trajectory prediction, we need a way to
obtain the relative positions of other vehicles and this is done 𝑝(𝒚|𝒇)~𝑁(𝒚|𝒇, 𝜎𝑁2 𝑰) (3)
through You Only Look Once (YOLO) [14]. We choose YOLO
Using the definition of the Gaussian Process in (1), we can
object detection algorithm because YOLO is the state-of-the-art
derive the prior to be as follows:
algorithm which is able to achieve real time detection with high
accuracy. 𝑝(𝒇|𝑿)~𝑁(𝒇|𝝁(𝑿), 𝑲(𝑿, 𝑿)) (4)
In equation (4), vectors containing function values f and
mean function 𝝁(𝑿) are assessed at each training input with 𝒇 =
Car Detected! [𝑓(𝑥1 ), … , 𝑓(𝑥𝑇 ))]𝑇 and 𝝁(𝑿) = [𝑚(𝑥1 ), … , 𝑚(𝑥𝑇 ))]𝑇
respectively. Lastly, 𝑲(𝑿, 𝑿), also known as a kernel, is a
covariance matrix which contains the covariance between all the
inputs.
Now that we have the prior and likelihood, we can now
predict new data by calculating the predictive posterior:
𝑝(𝒚∗ |𝑿∗ , 𝑿, 𝒚) = ∫ 𝑝(𝒚∗ |𝑿∗ , 𝒇, 𝑿)𝑝(𝒇|𝑿, 𝒚)𝑑 𝒇 (5)
In right side of the integral in (5), it can be derived from
equations (2)(3) and (4) and on the left side, it is the likelihood
Figure 2: YOLO Car detection of the new data 𝒚∗ over the function 𝒇. The left side of the
integral can be obtained by using known identities regarding
Based on the YOLO algorithm, we can obtain the coordinates Gaussian distribution as
of the top left and bottom right of bounding box which is used to
derive center point of the vehicle detected. Using this method 𝑝(𝒚∗ |𝑿∗ , 𝒇, 𝑿)~𝑁(𝑲(𝑿∗ , 𝑿)𝑲(𝑿∗ , 𝑿)−1 𝒇,
assumes that the bounding box for YOLO reflects the actual 𝑲(𝑿∗ , 𝑿∗ ) − 𝑲(𝑿∗ , 𝑿)𝑲(𝑿, 𝑿)−1 𝑲(𝑿, 𝑿∗ )) (6)
position of the car in (x,y) pixel coordinates although the YOLO
bounding box may deviated from the true center. In order to where 𝑲(𝑿∗ , 𝑿), 𝑲(𝑿, 𝑿∗ ) are covariance between training and
minimize this error, we will pass the collected data through a low new input data. Lastly, we can calculate the predictive posterior
pass filter to smoothen the time series data of (x,y) coordinates distribution using the integral (5) to obtain:
against time and with the filtered data, we can predict trajectory
more accurately with GPAR. 𝑝(𝒚∗ |𝑿∗ , 𝑿, 𝒚)~𝑁(𝒚∗ |𝝁∗ , 𝚺∗ ) (7)
Where, 𝝁∗= 𝑲(𝑿∗ , 𝑿)[ 𝑲(𝑿, 𝑿) + 𝜎𝑁2 𝑰]−1 𝒚
𝚺∗ = 𝑲(𝑿∗ , 𝑿∗ ) − 𝑲(𝑿∗ , 𝑿)[ 𝑲(𝑿, 𝑿) + 𝜎𝑁2 𝑰]−1 𝑲(𝑿, 𝑿∗ ) combination of various kernels. The method of combining
different kernels for Gaussian Process Regression is called
From equation (7), it is easy to understand that a Gaussian
Structural modelling as we are modelling the data based on
Process Regression is a non-parametric regression model as it
certain characteristic which can be described by the kernels. In
infers functions as ‘parameters’. Moreover, (7) shows that the
[16], forecasting of carbon dioxide emission was successful by
Gaussian Process Regression is heavily dependent on the
using the summation of Radial Basis Function kernel, periodic
kernels used and hence, it is important to understand the
kernel and linear kernel for the Gaussian Process Regression.
different kinds of kernel and how it affects the model.
This means that structural modelling should be able to predict
The kernel functions that will be discussed are the Radial our vehicle coordinates trajectory if we were to know the trend
Basis Function (RBF) kernel, linear kernel and periodic kernel and which kernel to use.
and the respective equations are as follows:
However, for our time series, while there may be trends over
(𝑥−𝑥 ′ )
2
a small period 𝑡𝑖𝑚𝑒 = 𝑡 to 𝑡𝑖𝑚𝑒 = 𝑡 + 𝑇, the trend is may not
𝑘𝑅𝐵𝐹 (𝑥, 𝑥 ′ ) = 𝜎 2 exp (− ) (8) be consistent throughout and structured modelling for trajectory
2𝜆2

Radial Basis Function kernel smoothens the regression as it prediction is not favorable as we must first identify the trends to
assumes that the underlying function is smooth and infinitely use for every small time period. Hence, we need to explore other
differentiable. By tuning the parameter λ, we can change the models and integrate it with Gaussian process to make it more
smoothness of the graph. dynamic.

𝜋(𝑥−𝑥′ ) C. Gaussian Process Auto-Regression (GPAR)

𝑠𝑖𝑛2 ( )
𝑘𝑝𝑒𝑟𝑖𝑜𝑑𝑖𝑐 (𝑥, 𝑥 ′) 2
= 𝜎 exp (−
𝜌
) (9) Instead of identifying the trends for structure modelling, we
2𝜆2 can combine Gaussian Process Regression with Auto-
The periodic kernel is also known as the exp-sine-squared Regression method. For Gaussian Process Auto-Regression, the
kernel and it is used when the time series we want to predict has model will now take past data of (x, y) coordinates as an input
a periodic trend. vector to produce our output. The input and output vectors are
as follows for a 𝜌-th order auto-regression, and a ℎ-step ahead
𝑘𝐿𝑖𝑛𝑒𝑎𝑟 (𝑥, 𝑥 ′ ) = 𝜎 2 𝑥𝑥′ (10) forecast:
The linear kernel is used when the data show a linear trend 𝑦1 ⋯ 𝑦𝑝 𝑦𝑝+ℎ
such as a linearly increasing or decreasing graph. 𝑥=[ ⋮ ⋱ ⋮ ] , 𝑦=[ ⋮ ] (11)
The most commonly used kernel is the Radial Basis because 𝑦𝑇−ℎ−𝑝+1 ⋯ 𝑦𝑇−ℎ 𝑦𝑇
Gaussian Process Regression with RBF kernel give a smooth According to [9], there are 2 methods of forecasting and they
preidiction for interpolation of stationary time series data. A are the recursive and direct method. In this section, we will
Stationary time series is a time series data set which properties discuss the 2 methods and why we chose to use direct method.
doesn’t change with time which does not applies to trajectory
prediction of vehicle coordinates. However, vehicle coordinates Recursive Forecasting
data is not stationary and hence, it is difficult for GPR with RBF For the recursive method, as the name suggest, first predicts
to predict future vehicle coordinates accurately. 𝑦𝑇+1 and using this value of 𝑦𝑇+1 as the input to recursively
predict the next step until we predict 𝑦𝑇+ℎ . With the test input
𝒙∗ = [𝑦𝑇−𝑝 , 𝑦𝑇−𝑝+1 … , 𝑦𝑇 ]𝑇 , we can use Gaussian Process
Auto-Regression to estimate 𝑦𝑇+1 .
𝑦𝑇+1 ~ 𝑁(𝑦𝑇+1 |𝜇1 , 𝜎12 ) (12)
The next iteration will be using the estimated 𝑦𝑇+1 mean
value of 𝜇1 inside our new test input: 𝒙∗ =
[𝑦𝑇−𝑝+1 , 𝑦𝑇−𝑝+2 … , 𝑦𝑇, 𝜇1 ]𝑇 to predict 𝑦𝑇+2 .
𝑦𝑇+2 ~ 𝑁(𝑦𝑇+2 |𝜇2 , 𝜎22 ) (13)
These steps will be repeated recursively to predict h-step-
ahead forecast, 𝑦𝑇+ℎ . While this method is straightforward, this
model suffers from error propagation, because the forecast value
is computed from the previously predicted value which
decreases the accuracy with every step.
Figure 3: Trajectory Prediction of x coordinates against time
step t with GPR on RBF kernel. Direct Forecasting

As shown from Figure 3, using Gaussian Process Regression On the other hand, direct strategy is computing the ℎ-step
with Radial Basis Function kernel is not useful because our data ahead prediction directly and every (T+1), ( T+2), …., (T+h)
is not stationary and extrapolation of such data will forecast steps will be computed individually. This method is generally
values close to the mean 𝑚(𝑥) and since Gaussian Process more expensive than recursive strategy as it has to train a
Regression typically assume mean of zero, the x coordinates different model for every (T+1), ( T+2), …., (T+h) steps. The
predicted return to zero after several steps. model for h-step ahead is as follows:

B. Structural modeling 𝑦̂𝑇+ℎ = 𝑓ℎ (𝑦𝑇−𝑝 , 𝑦𝑇−𝑝+1 … , 𝑦𝑇 ) (14)

In order to improve extrapolation predictions, a more The advantage for direct strategy is that the error does not
complex covariance function 𝑘̂(𝑡, 𝑡 ′ ) can be applied to Gaussian propagate, and this is usually the justification on why direct
Process Regression and the new kernel is simply the strategy is better than recursive method. But the disadvantage is
that it trains the (T+1), ( T+2), …., (T+h) predicted values
separately and they are independent from each other which is
not true for most cases.
V. EXPERIMENTAL RESULTS
Our initial experiments are conducted using trajectory
detected from a 10 second video with 29.97 frames per second.
We recorded (x, y) coordinates of the bounding box into a list,
filter it through a low pass filter and this will be our input data
for trajectory prediction. We used various models which
includes Gaussian Process Regression, Gaussian Process Auto-
Regression (GPAR), Moving Average and lastly, Kalman Filter
and in this section, we will evaluate the various models using
the mean-squared-error (MSE).
A. Structure Modelling with Gaussian Process Regression
Trajectory prediction results Figure 5: GPAR 20-steps prediction.
As seen in figure 5, when autoregression model is integrated
with the Gaussian Process, the predicted values are able to
follow the trajectory fairly well.
1) Moving average Trajectory prediction results

Figure 4: Trajectory prediction of x coordinates using structure

modelling based on Radial Basis Function Kernel + Linear
Kernel.
As mentioned previously in Section IV part B, structural
model can be used for forecasting but as seen in figure 4,
structural modelling still has its limitations as it is heavily
dependent on the kernels hyperparameters. The gradient of the
predicted line is heavily dependent on the linear kernel
parameters.
B. Gaussian Process Auto-Regression Trajectory prediction
results

Figure 6: Moving Average Forecasting of velocities with x

(top), y (bottom) coordinates.
For Moving Average forecasting, we used the moving
average of velocity values because we know that the moving
average will converge to a mean value and this ensures that
the coordinate values of x and y will follow a constant
velocity for trajectory forecasting. However, this is not
enough as there will be accelerations as well.
C. Kalman Filter Trajectory prediction results
Since conventional Kalman filter is able to predict 1-step
ahead, we need to devise a way to predict h-steps ahead. In
Kalman filter, there are 2 update steps specifically the
measurement update and state transition update. Since there
are no future values for measurement update, we can use the
 transition matrix to predict h-steps ahead. In this section, we
will be using the constant acceleration model of the Kalman
filter.

Figure 8: Kalman Filter for Smoothing. Red line is the filtered

data and blue line is the raw data without filtering.
Figure 7 and 8 shows the use of a Kalman filter to do h-steps
ahead prediction and Kalman filter for smoothing curve. While
Kalman filter does a great job at smoothing noisy data, it is
unable to predict trajectory very well mainly because it only
relies on the transition matrix for prediction. Since transition
matrix are based on kinematics model, it fails to consider the
everchanging dynamics of road conditions. Moreover, Kalman
filter will also produce a lag which is evident in both figure 7
and 8. In figure 7, there is a gap between the 1-step ahead
predicted value and actual value and in figure 8, the curve of the
Kalman filter follows after the actual values.
D. Comparison of GPAR, moving average and Kalman Filter
using Mean Square Error
GPAR Moving Kalman
Average Filter
MSE, x 57 957 1743
MSE, y 665 4563 256
Figure 7: Kalman Filter Forecasting 8 steps. First 10 frames are Sum 722 5520 1999
filtered data and subsequent frames are predicted data.
Table 1: Comparison using Mean Square Error (pixel distance
square)
As seen from the table above, it is concluded that GPAR
works best for trajectory prediction with Kalman filter as the
second and moving average as the last. This is mainly because
autoregression uses p-lag values for its future predictions
through a covariance matrix while Kalman filter only uses its
previous value for its prediction. Although moving average also
uses p-lag values, it weights the values equally while Kalman
filter and GPAR doesn’t and hence, it has the largest Mean-
squared-error.

VI. CONCLUSION
Time Series forecasting has been a focus for many years, but
the effectiveness of the algorithm really depends on the nature
of the time series data. While there has been much research into
trajectory prediction for autonomous vehicles and Advance
Driver Assistance System (ADAS) it mainly requires many
sensors to accurately detect the ego’s vehicle position, velocity,
yaw rate and even more sensors are used to detect other vehicles
relative position and velocity. This paper proposes a method to
predict trajectory of other vehicles based on a singular camera
platform using Gaussian Process Auto-Regression which is not
limited to physics and kinematics model. This paper has proven
that using GPAR gets better results than existing forecasting
methods such as Kalman Filter with constant acceleration
model, moving average and Gaussian Process Regression.

REFERENCES
[1] Y. Global status report on road safety 2015, World Health Organization,
2015.
[2] S. Satriawan, C. Machbub and E. M. I. Hidayat, "Comparison of prediction
methods for moving objects in 3D coordinates using Kalman Filter and
Least Square," 2016 6th International Conference on System Engineering
and Technology (ICSET), Bandung, 2016, pp. 128-131.
[3] X. Mu, J. Che, T. Hu and Z. Wang, "A video object tracking algorithm
combined Kalman filter and adaptive least squares under occlusion," 2016
9th International Congress on Image and Signal Processing, BioMedical
Engineering and Informatics (CISP-BMEI), Datong, 2016, pp. 6-10.
[4] . J. Schulz, C. Hubmann, J. Löchner and D. Burschka, "Multiple Model
Unscented Kalman Filtering in Dynamic Bayesian Networks for Intention
Estimation and Trajectory Prediction," 2018 21st International Conference
on Intelligent Transportation Systems (ITSC), Maui, HI, 2018, pp. 1467-
1474.
[5] Q. Lim, Y. He and U. Tan, "Real-Time Forward Collision Warning
System Using Nested Kalman Filter for Monocular Camera," 2018 IEEE
International Conference on Robotics and Biomimetics (ROBIO), Kuala
Lumpur, Malaysia, 2018, pp. 868-873.
[6] M. Yu, W. Zhou, B. Wang and J. Jin, "The short-term forecasting of wind
speed based on EMD and ARMA," 2017 12th IEEE Conference on
Industrial Electronics and Applications (ICIEA), Siem Reap, 2017, pp.
495-498.
[7] J. Rodrigues and A. Deshpande, "Prediction of Rainfall for all the States
of India Using Auto-Regressive Integrated Moving Average Model and
Multiple Linear Regression," 2017 International Conference on
Computing, Communication, Control and Automation (ICCUBEA), Pune,
2017, pp. 1-4.
[8] S. A. Goli, B. H. Far and A. O. Fapojuwo, "Vehicle Trajectory Prediction
with Gaussian Process Regression in Connected Vehicle
Environment$\star$," 2018 IEEE Intelligent Vehicles Symposium (IV),
Changshu, 2018, pp. 550-555.
[9] E. J. Heravi and S. Khanmohammadi, "Long Term Trajectory Prediction
of Moving Objects Using Gaussian Process," 2011 First International
Conference on Robot, Vision and Signal Processing, Kaohsiung, 2011, pp.
228-232.
[10] K. Saleh, M. Hossny and S. Nahavandi, "Intent Prediction of Pedestrians
via Motion Trajectories Using Stacked Recurrent Neural Networks," in
IEEE Transactions on Intelligent Vehicles, vol. 3, no. 4, pp. 414-424, Dec.
2018.
[11] S. Dai, L. Li and Z. Li, "Modeling Vehicle Interactions via Modified
LSTM Models for Trajectory Prediction," in IEEE Access, vol. 7, pp.
38287-38296, 2019.
[12] L. Xin, P. Wang, C. Chan, J. Chen, S. E. Li and B. Cheng, "Intention-
aware Long Horizon Trajectory Prediction of Surrounding Vehicles using
Dual LSTM Networks," 2018 21st International Conference on Intelligent
Transportation Systems (ITSC), Maui, HI, 2018, pp. 1441-1446.
[13] S. H. Park, B. Kim, C. M. Kang, C. C. Chung and J. W. Choi, "Sequence-
to-Sequence Prediction of Vehicle Trajectory via LSTM Encoder-Decoder
Architecture," 2018 IEEE Intelligent Vehicles Symposium (IV),
Changshu, 2018, pp. 1672-1678.
[14] J. Redmon, S. Divvala, R. Girshick and A. Farhadi, "You Only Look
Once: Unified, Real-Time Object Detection," 2016 IEEE Conference on
Computer Vision and Pattern Recognition (CVPR), Las Vegas, NV, 2016,
pp. 779-788.
[15] Carl Edward Rasmussen and Christopher K. I. Williams. Gaussian
Processes for Machine Learning. University Press Group Limited, January
2006. ISBN 978-0-262-18253-9.
[16] Bagas Abisena Swastanto (2016), Gaussian Process Regression for Long-
Term Time Series Forecasting (Master Thesis). Retrieved from
https://repository.tudelft.nl/islandora/object/uuid:7f3916c9-795c-403f-
bec8-5eb6caa3e398/datastream/OBJ/download

View publication stats