End-to-end Learning Approach for Autonomous Driving: A

Convolutional Neural Network Model

Yaqin Wang, Dongfang Liu, Hyewon Jeon, Zhiwei Chu and Eric T. Matson

Department of Computer and Information Technology, Purdue University, U.S.A.

Keywords:

Autonomous Driving, AI, Convolutional Neural Network, End-to-end Approach.

Abstract:

End-to-end approach is one of the frequently used approaches for the autonomous driving system. In this

study, we adopt the end-to-end approach because this approach has been approved to lead to a distinguished

performance with a simpler system. We build a convolutional neural network (CNN) to map raw pixels from

cameras of three different angles and to generate steering commands to drive a car in the Udacity simulator.

Our proposed model has a promising result, which is more accurate and has lower loss rate comparing to

previous models.

1 INTRODUCTION

In 1989, (Pomerleua, 1989) designed a simple struc-

tured neural network with three fully-connected lay-

ers to steer a car on the CMU campus. This was an

initial research to demonstrate that machine learning

theory can be implemented to generate commands to

guide a car. However, until now, no autonomous cars

have been designed which meet the full autonomous

driving (R

odel et al., 2014). As the most important

component in autonomous driving, different models

based on different approaches have been used to gen-

erate appropriate commands to steer the car. Progress

of autonomous driving is due to the optimization of

the model design.

Currently, the mediated perception approach and

end-to-end approach are two frequently used ap-

proaches for autonomous driving system. Mediated

perception approach is widely used in the car industry

which requires a large assistive system (lidar, radar,

sensor, and central computing systems) for detection,

prediction, and decision making (Chen et al., 2015).

On the contrary, the end-to-end approach is a much

lighter system and utilizes a distinct theory to steer

a car (Bojarski et al., 2016)(Xu et al., 2017)(Codev-

illa et al., 2018)(Pfeiffer et al., 2017). This approach

focuses on building an artiﬁcial intelligence model

which emulates humans driving. For human, driv-

ing on the road, steering commands are constantly

made based on the instant processing of images (Bo-

jarski et al., 2016)(Xu et al., 2017)(Codevilla et al.,

2018)(Pfeiffer et al., 2017). For an end-to-end learn-

ing approach, the model is normally trained by learn-

ing the humans actions on the road (Bojarski et al.,

2016)(Xu et al., 2017)(Codevilla et al., 2018)(Pfeif-

fer et al., 2017). Here, one end is reading the input

image and another end is to respond accordingly to

road conditions. Using end-to-end learning approach

is easy to maneuver and cost-effective compared to

mediated perception approach.

Prior research has explored the end-to-end learn-

ing approach for autonomous driving. A pioneering

study is from (Muller et al., 2006) who utilized an

end-to-end learning system to design an autonomous

car. (Muller et al., 2006) used a six-layer convo-

lutional neural network for image processing. This

CNN model can process the raw pixels and com-

mand steering angels based on the input image. In

addition, (Chen et al., 2015) employed the Caffe

framework to develop a CNN model to process the

dataset from TORCS (The Open Racing Car Sim-

ulator). The model has a higher computing speed

compared to the model from (Muller et al., 2006)

when making commands for steering actions. An-

other distinguished end-to-end learning model is from

(Su et al., 2017). To improve the generalization of

their model, (Su et al., 2017) employed a large dataset

for training, which included different road conditions.

The outcomes of this research obtained a high ac-

curacy pertaining to steering angels compared to its

previous counterparts. CNN models have also been

widely used in object detection and classiﬁcation in

autonomous driving (Ren et al., 2015)(Girshick et al.,

2014).

Wang, Y., Liu, D., Jeon, H., Chu, Z. and Matson, E.

End-to-end Learning Approach for Autonomous Driving: A Convolutional Neural Network Model.

DOI: 10.5220/0007575908330839

In Proceedings of the 11th International Conference on Agents and Artiﬁcial Intelligence (ICAART 2019), pages 833-839

ISBN: 978-989-758-350-6

833

Aside from the seminal studies discussed above,

a widely recognized research is from (Bojarski et al.,

2016). The researchers proposed an innovative CNN

model based on the end-to-end learning approach.

The strength of this research is that the CNN model

has a much simpler structure compared to the previ-

ous studies. It only has nine layers of networks with

ﬁve convolutional layers and four fully-connected

layers. Hence, the computing speed for this model

is faster which can process 30 frame per second. In

addition, this model does not require a large dataset

for training. The total training dataset for this model

only included 100 hours of video clips of different

road conditions.

Although the previous research has obtained var-

ious achievements for end-to-end learning approach,

we identiﬁed a number of problems in the process of

survey studies of previous research. First, the ma-

jority of the previous end-to-end learning approach is

based on the CNN model for image processing. CNN

is an effective neural framework to process image, but

it is also subject to the problems of overﬁtting and

vanishing gradient in the training phase (Srivastava

et al., 2014)(Ma et al., 2017)(Klein et al., 2018)(Su

et al., 2017). Second, a desirable outcome of the

end-to-end learning approach is that the autonomous

car can emulate the humans driving and keep on the

road with limited human interventions. However,

most autonomous cars based on the end-to-end learn-

ing approach need a large number of human inter-

ventions in previous research (Pomerleua, 1989)(Xu

et al., 2017)(Muller et al., 2006)(Chen et al., 2015).

The contributions of this study are in three folds.

First, we design a CNN model based on end-to-end

learning approaches which has better generalization

than that of (Bojarski et al., 2016). We will discuss

more details in the proposed model section pertaining

to how we improve the noise robustness of our model.

This goal is evaluated by testing the generalization of

our proposed model. We evaluate the loss rate and ac-

curacy of our model during the training and address

the potential problem of overﬁtting and vanish gra-

dient. Second, our proposed CNN model based on

end-to-end learning approach can steer a car longer

than the reported time from (Bojarski et al., 2016).

We compare the results and report them in the result

and discussion section.

The rest of this report is organized as follows: In

section 2, the required materials and the proposed

methods will be introduced; In section 3, we will

present the results and discussions about our project;

In section 4, we will conclude the progress of our on-

going project and offer some insights for the future

work.

2 METHODOLOGY

2.1 Simulator

There is a large array of autonomous driving simu-

lators available on GitHub, as seen in Table 1. The

selection criteria for our project are that: 1. The sim-

ulator does not have a high requirement for graphics

device; 2. The simulator can easily collect image data

and generate CSV ﬁle which includes labeling infor-

mation for the collected images; 3. The simulator can

be installed and operated on Mac OS X; 4. The simu-

lator has an autonomous driving mode to test the au-

tonomous model.

We ﬁnally decide to use Udacitys Self-Driving

simulator, because it ﬁts well with our selection stan-

dards. This simulator platform has high graphics re-

quirements and it can be easily installed on Mac OS. It

has training mode and autonomous mode. In training

mode, the simulator can record images and generate

a CSV ﬁle simultaneously when we drive a car in its

environment. In autonomous mode, we can utilize the

designed model to programmatically navigate the car

in the environment.

2.2 Data Collection

The data collection is iterative and has two phases.

In the ﬁrst phase, we focused on practicing the driv-

ing skills. After the Udacity Self-Driving simulator is

installed, we started to use training mode to practice

the driving skills. We constantly evaluated our per-

formance of driving in the simulator and improve the

driving accuracy pertaining to keeping the car in the

middle of the lane. The ﬁrst phase lasted about one

month before we started collecting data for training.

In the second phase, we made strict rubrics for the

collection process which determine the usability of

the collected data. The usable data only includes im-

ages that the car is running inside of two-lane marks

and the car does not suffer from drastic steering an-

gle changes. Data which fails to meet the rubrics is

not included in the ﬁnal dataset for training. Figure 1

demonstrates the real data collection process, which

simultaneously recorded the left, center, and right pic-

ture of a running car in training mode. A CSV ﬁle is

also automatically generated while we run the car in

Udacity. A sample CSV ﬁle includes information of

labels of center, left, and right pictures as well as a

cars current steering angle, accelerating, reserve, and

speed, as seen in Figure 2. The second phase of data

collection lasted around one month, we run a car for

approximately sixty hours on Udacity and we ﬁnally

decided to include 83,424 pictures for model training.

ICAART 2019 - 11th International Conference on Agents and Artiﬁcial Intelligence

834

Table 1: Summary of the autonomous driving simulators.

Name Graphic require-

ment

Data collection OS requirements (Mainly sup-

port)

Autonomous driving

mode

Udacity low easy Window, Mac OS, Linux Yes

TORCS medium complex Linux Yes

Euro Truck Simulator 2 high complex Window, Mac OS, Linux Yes

Gazebo medium complex Linux Yes

CARLA high complex Window, Linux Yes

a b

Figure 1: Left (a), center (b), and right (c) camera of a running car.

2.3 Method

In this study, we build a convolutional neural network

(CNN) to map raw pixels from cameras of three dif-

ferent angles which generate steering and acceleration

commands to drive a car in the Udacity simulator. The

overall framework for this research is demonstrated

in Figure 3. An end-to-end approach is employed be-

cause this approach has been approved to lead to a dis-

tinguished performance with a simpler system (Bo-

jarski et al., 2016). A distinguished performance can

be achieved because the end-to-end learning system is

trying to emulate humans action, instead of learning

human-selected criteria, such as lane detection or lane

keeping. In addition, a simpler solution is deﬁned

by the fact that an end-to-end system only requires a

small number of datasets for training (Bojarski et al.,

2016). Theoretically, if the training dataset only in-

cludes the actions we want the model to learn, an end-

to-end learning system is able to achieve a productive

outcome with less effort for training. With the road

images associated with steering angles and accelera-

tion as the training inputs, our system automatically

learns internal representations of the imperative pro-

cessing steps pertaining to making proper commands

for driving. We do not need to explicitly train our

model to detect road features, for instance, the outline

of roads or background objects. Our model is able to

make command to change steering angles and accel-

eration when encountering similar road condition.

2.4 The Proposed Model

In this study, we rely on the framework from (Bojarski

et al., 2016) to build our model. The study from (Bo-

jarski et al., 2016) is the widely recognized end-to-

end learning approach which designed a CNN model

for autonomous driving in recent years. However, like

all CNN based model, (Bojarski et al., 2016) is sub-

ject to the problem of overﬁtting and vanishing gra-

dient (Srivastava et al., 2014)(Ma et al., 2017)(Klein

et al., 2018)(Su et al., 2017). Also, the loss rate and

accuracy for (Bojarski et al., 2016) can be further im-

proved.

To address the issue of overﬁtting and eventually

improve the loss rate and accuracy, we purposefully

add dropout layers to our model. Dropout is a pow-

erful regularization method for Multi-Layer Percep-

trons (MLPs) trained by Stochastic Gradient Descent

(SGD) (Szegedy et al., 2016)(Clevert et al., 2015)(Gal

and Ghahramani, 2016). During the forward propaga-

tion step of the SGD, the dropout layers consist in set-

ting zero to a random subset of the hidden activation.

Accordingly, the backward propagation in the SGD

is also modiﬁed. By this mean, dropout is able to

prevent the co-adaptation of learned hidden features

which is a source of overﬁtting. Eventually, the net-

work becomes less sensitive to the speciﬁc weights of

neurons and has a better capacity of generalization.

Figure 3 demonstrated the overview of the pro-

posed model in the study. For the ﬁrst ﬁve layers,

we use convolutional layers to capture the key ﬁg-

ures of the input data. The part of the structure is

derived from (Bojarski et al., 2016), which has been

approved effective to capture road features. After the

ﬁrst ﬁve layers of iterative processing, the feature im-

ages for input data is well-learned by the system. The

output of convolutional layers is connected to four

fully-connected layers. At this stage, we try to add a

dropout layer between fully-connected layers because

a number of research (Szegedy et al., 2015)(Szegedy

et al., 2016)(Clevert et al., 2015) suggested that

End-to-end Learning Approach for Autonomous Driving: A Convolutional Neural Network Model

835

Figure 2: A Sample CSV ﬁle.

Figure 3: The Overall Framework.

dropout layer should be used between fully-connected

layers and not in convolutional layers. However, there

is a debate pertaining to where a dropout layer should

be added (Chen et al., 2018)(Clevert et al., 2015)(Gal

and Ghahramani, 2016)(Bluche et al., 2015). One

widely accepted solution is to add dropout layers

in different locations in the neural network and test

its overall performance (Chen et al., 2018)(Clev-

ert et al., 2015)(Gal and Ghahramani, 2016)(Bluche

et al., 2015). Following this rule, we purposefully

add the dropout layer in different locations between

the fully-connected layers and test the performance

for the entire network. In addition, considering the

fact that the dataset is relatively small, data augment

function is also carefully designed in order to create a

new data and prevent overﬁtting. The augment func-

tion we design include ﬂipping (both vertically and

horizontally), translating (moving along either of the

x, y axis), cropping, and rotating.

In addition, when designing our model, we care-

fully compare four commonly used activation func-

tions such as ELU, LReLU, ReLU, and SReLU, in

order to address the problem of vanishing gradient.

After the comparison across the four activation func-

tions, we ﬁnally use ELU as our activation function

for our model for two main reasons. First, when pro-

cessing positive values, all of the four activation func-

tions have good performance for solving the problem

of vanish gradient. However, if the processed values

are negative, RelU will cause bias to the next neural

network layer because the output of RelU is zero. The

bias will be added cross later layers and amplify the

effect of bias weight. Second, although the outputs of

LReLU and PReLU is not zero when processing nega-

tive inputs, the two activation functions cannot main-

tain noise robustness when the input value is small.

On the contrast, ELU has soft saturation character-

istics, so when the input is a small negative value, it

improves the noise robustness and has a desirable per-

formance to address the problem of vanish gradient

(Szegedy et al., 2015)(Szegedy et al., 2016).

3 RESULTS AND DISCUSSIONS

3.1 Training Results

In the experiment, we utilize 32 as our batch size,

20,000 as our maximum training steps, and 15 as our

epochs for each training circle. The trained model

is automatically saved after the training. We dupli-

cate the model named NVIDIA and use the same

ICAART 2019 - 11th International Conference on Agents and Artiﬁcial Intelligence

836

Table 2: The Results of Validation Loss Rate and Accuracy for Different Models. NVIDIA model is the model from (Bojarski

et al., 2016); Model 1 is the proposed model with dropout layer after the ﬁrst fully-connected layer; Model 2 is the proposed

model with dropout layer after the second fully-connected layer; Model 3 is the proposed model with a dropout layer after the

third fully-connected layer.

Epoch 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15

NVIDIA Loss rate 0.025 0.025 0.025 0.024 0.022 0.022 0.023 0.022 0.024 0.024 0.025 0.023 0.023 0.023 0.023

Accuracy 0.776 0.78 0.779 0.779 0.784 0.776 0.778 0.783 0.779 0.784 0.783 0.776 0.782 0.776 0.779

Model 1 Loss rate 0.026 0.026 0.026 0.026 0.027 0.028 0.029 0.028 0.028 0.028 0.029 0.028 0.027 0.027 0.027

Accuracy 0.76 0.76 0.764 0.765 0.765 0.774 0.775 0.76 0.76 0.77 0.75 0.76 0.76 0.76 0.76

Model 2 Loss rate 0.027 0.027 0.028 0.027 0.026 0.025 0.026 0.026 0.027 0.026 0.027 0.026 0.026 0.026 0.026

Accuracy 0.766 0.762 0.75 0.74 0.765 0.775 0.775 0.774 0.773 0.776 0.776 0.774 0.774 0.774 0.774

Model 3 Loss rate 0.025 0.025 0.025 0.024 0.022 0.022 0.022 0.022 0.022 0.021 0.021 0.021 0.021 0.021 0.022

Accuracy 0.772 0.778 0.779 0.776 0.781 0.783 0.773 0.778 0.776 0.777 0.775 0.78 0.778 0.78 0.781

dataset to train our models and the NVIDIA model.

We also add the dropout layer at different location

across fully-connected layers. The results for train-

ing dataset are demonstrated in Figure 6. Based on

the results, we can see that proposed models with

dropout layer have better training outcomes compared

to NVIDIA model. The ﬁnal loss rate for models

with the dropout layer is smaller than that of (Bojarski

et al., 2016), as seen in Table 2. Also, the ﬁnal accu-

racy for models with dropout layer is also higher than

that of (Bojarski et al., 2016), as seen in Table 2. The

proposed model 3 has best performance pertaining to

having lowest loss rate and highest accuracy in train-

ing dataset among all presented models. Model 3 is

the proposed model with dropout layer after the third

fully-connected layer.

Next, we further compare the proposed model 3

with the NVIDIA model in terms of the changing of

the loss rate in training and validation process, as seen

in Figure 5. The loss rate for model 3 in both train-

ing and validation continuously decreases. In con-

trast, the validation loss rate for the NVIDIA model

starts to increase at epoch 12 which is an indicator

for overﬁtting. Based on the results, adding dropout

layer effectively addresses the issue of overﬁtting and

improves the noise robustness of the original model.

The main contribution of this study is to build a

model that could surpass the performances of previ-

ous studies.

3.2 Generalization and Lane Keeping

Time

We purposefully evaluate the generalization of our

model. The Udacity simulator has two road tracks

with totally different road conditions. One road track

is a lake road and the other is a mountain road. Lake

road is very ﬂat and has no steep curves. On the con-

trary, the mountain road has challenging road condi-

tions with a lot of steep curves. We use the mountain

road dataset to train our model and test the model to

Figure 4: The Comparison of ELU, LReLU, ReLU, and

SReLU.

run a car in the lake road track. Since a car running on

a mountain road needs frequent turns of steering an-

gle, if the trained model simply clones the behaviors

in the dataset, the generalization of the model could

be compromised because running on a ﬂat road does

not need a large degree of steering angle. However,

the result from our experiment demonstrates that our

model has a desirable performance and it runs a car

on lake track.

In addition, we test the lane keeping time for our

model. We test our model 30 times on the lake road.

When running on the lake road track, our model can

End-to-end Learning Approach for Autonomous Driving: A Convolutional Neural Network Model

837

Figure 5: The Overall Framework.

Figure 6: Frequency Distribution of the NVIDIA Model and Model 3.

steer the car on the road and keep the car on the road

for more than 617.3 seconds average, as seen in Table

3. In addition, we use the NVIDIA model to steer a

car on the lake road track. The NVIDIA model can

keep the car in the road without running off the road

for 453.7 seconds average, as seen in Table 3. Since

we test the two model by using the same car and same

road track, we use paired t-test to evaluate the perfor-

mance difference between two models. We ﬁrst check

the assumption of the paired t-test and ensure that the

collected data is normal distributioned, as seen in Fig-

ure 6. Then we run the paired t-test for the data. The

result is signiﬁcant (p<0.0001) and Model 3 with the

dropout layer is better than NVIDIA model without

the dropout layer because of longer lane keeping time,

as seen in Table 4.

4 CONCLUSION

In this study, we design a CNN model based on an

end-to-end learning approach and address the prob-

lem of overﬁtting and vanish gradient. Our proposed

model outperforms the previous study which we rely

on as our model framework. In addition, we evaluate

the generalization of our proposed model. We use a

speciﬁc road condition to train our model and test it

on a different road condition. Finally, we record how

much time the car can remain in the lane without hu-

man intervention and the result indicates that the car

can stay on the road longer than the NVIDIA model.

Limitations for the project are a few. First, the ac-

tion for the end-to-end learning approach is deﬁned

as steering angles and change of speed rate in our

study. However, actions for real driving scenario is

more complicated and should include reverse, brake,

and more. In addition, we did not consider obstacles,

such as pedestrians and other cars, in our research.

Future research can add object detection functions in

the model to simulate real road conditions.

ICAART 2019 - 11th International Conference on Agents and Artiﬁcial Intelligence

838

Table 3: Lane Keep Time: The report time is in second and

rounded up for keeping two decimals.

Test time Model 3 NVIDIA Model

1 582.3 450.6

2 624.5 472.4

3 541.7 442.3

4 693.1 443.7

5 599.2 458.3

6 612.4 451.4

7 643.8 454.2

8 605.9 451.8

9 564.1 484.9

10 642.4 438.4

11 656.4 460.4

12 604.5 464.1

13 587.1 434.8

14 627.2 469.9

15 623.2 457.5

16 674.4 428.7

17 675.7 449.4

18 611.8 456.3

19 604.2 461.6

20 572.8 443.4

Average 617.3 453.7

Table 4: Paired t-test for Model 3 and NVIDIA Model.

Variable Mean Statistic p-value

Model 3 617.3 16.63 8.84e-11<0.0001

NVIDIA Model 453.7

REFERENCES

Bluche, T., Kermorvant, C., and Louradour, J. (2015).

Where to apply dropout in recurrent neural networks

for handwriting recognition? In Document Analy-

sis and Recognition (ICDAR), 2015 13th International

Conference on, pages 681–685. IEEE.

Bojarski, M., Del Testa, D., Dworakowski, D., Firner, B.,

Flepp, B., Goyal, P., Jackel, L. D., Monfort, M.,

Muller, U., Zhang, J., et al. (2016). End to end learn-

ing for self-driving cars. arXiv, Cornell University:1

604.07316.

Chen, C., Seff, A., Kornhauser, A., and Xiao, J. (2015).

Deepdriving: Learning affordance for direct percep-

tion in autonomous driving. pages 2722 – 2730.

Chen, T., Lu, S., and Fan, J. (2018). S-cnn: Subcategory-

aware convolutional networks for object detection.

IEEE transactions on pattern analysis and machine

intelligence, 40(10):2522–2528.

Clevert, D.-A., Unterthiner, T., and Hochreiter, S. (2015).

Fast and accurate deep network learning by exponen-

tial linear units (elus). arXiv, Cornell University:

1511.07289.

Codevilla, F., Miiller, M., L

opez, A., Koltun, V., and Doso-

vitskiy, A. (2018). End-to-end driving via conditional

imitation learning. In 2018 IEEE International Con-

ference on Robotics and Automation (ICRA), pages 1–

9. IEEE.

Gal, Y. and Ghahramani, Z. (2016). A theoretically

grounded application of dropout in recurrent neural

networks. In Advances in neural information process-

ing systems, pages 1019–1027.

Girshick, R., Donahue, J., Darrell, T., and Malik, J. (2014).

Rich feature hierarchies for accurate object detec-

tion and semantic segmentation. In Proceedings of

the IEEE conference on computer vision and pattern

recognition, pages 580–587.

Klein, A., Christiansen, E., Murphy, K., and Hutter, F.

(2018). Towards reproducible neural architecture and

hyperparameter search.

Ma, C., Zhu, Z., Ye, J., Yang, J., Pei, J., Xu, S., Zhou, R.,

Yu, C., Mo, F., Wen, B., et al. (2017). Deeprt: deep

learning for peptide retention time prediction in pro-

teomics. arXiv, Cornell University: 1705.05368.

Muller, U., Ben, J., Cosatto, E., Flepp, B., and Cun, Y. L.

(2006). Off-road obstacle avoidance through end-to-

end learning. In Advances in neural information pro-

cessing systems, pages 739–746.

Pfeiffer, M., Schaeuble, M., Nieto, J., Siegwart, R., and

Cadena, C. (2017). From perception to decision: A

data-driven approach to end-to-end motion planning

for autonomous ground robots. In 2017 ieee interna-

tional conference on robotics and automation (icra),

pages 1527–1533. IEEE.

Pomerleua, D. (1989). Alvinn: an autonomous land vehicle

in a neural network. Advances in neural information

processing systems, pages 305–313.

Ren, S., He, K., Girshick, R., and Sun, J. (2015). Faster

r-cnn: Towards real-time object detection with region

proposal networks. In Advances in neural information

processing systems, pages 91–99.

odel, C., Stadler, S., Meschtscherjakov, A., and Tscheligi,

M. (2014). Towards autonomous cars: the effect of

autonomy levels on acceptance and user experience.

Proceedings of the 6th International Conference on

Automotive User Interfaces and Interactive Vehicular

Applications, pages 1–8.

Srivastava, N., Hinton, G., Krizhevsky, A., Sutskever, I.,

and Salakhutdinov, R. (2014). Dropout: a simple way

to prevent neural networks from overﬁtting. The Jour-

nal of Machine Learning Research, 15(1):1929–1958.

Su, W., Chen, L., Wu, M., Zhou, M., Liu, Z., and Cao, W.

(2017). Nesterov accelerated gradient descent-based

convolution neural network with dropout for facial ex-

pression recognition. In Control Conference (ASCC),

2017 11th Asian, pages 1063–1068. IEEE.

Szegedy, C., Liu, W., Jia, Y., Sermanet, P., Reed, S.,

Anguelov, D., Erhan, D., Vanhoucke, V., and Rabi-

novich, A. (2015). Going deeper with convolutions.

In Proceedings of the IEEE conference on computer

vision and pattern recognition, pages 1–9.

Szegedy, C., Vanhoucke, V., Ioffe, S., Shlens, J., and Wo-

jna, Z. (2016). Rethinking the inception architecture

for computer vision. In Proceedings of the IEEE con-

ference on computer vision and pattern recognition,

pages 2818–2826.

Xu, H., Gao, Y., Yu, F., and Darrell, T. (2017). End-to-

end learning of driving models from large-scale video

datasets. arXiv, Cornell University.

End-to-end Learning Approach for Autonomous Driving: A Convolutional Neural Network Model

839