基於由普林斯頓大學Chen et al. in Proc. IEEE Int. Conf. Comput. Vis., 2722-2730, 2015 [1] 之論文提出的「自駕車之直觀感知方法」，這篇論文提出一個更完善的直觀感知架構與車輛控制演算法。我們在自駕車模擬平台TORCS之下設計出新的控制程式，並建構一個資料庫由圖片以及相對應5個特徵值組成，這是由13個特徵值版本[1]改良的。由我建構出來的資料庫來進行AlexNet之卷積神經網路(convolutional neural network) 迴歸訓練與執行自駕車模擬測試。我也採用不同的網路架構GoogLeNet進行訓練，並與AlexNet比較結果。實驗結果顯示兩者的training loss皆收斂相當完美，而且實際跑模擬顯示CNN自駕車能夠在未經模型學習過的賽道成功進行自動駕駛。

Based on the direct perception approach of autonomous driving proposed by Chen et al. in Proc. IEEE Int. Conf. Comput. Vis., 2722-2730, 2015 [1], we propose a more general direct perception framework and control algorithm in this thesis. We design a new controller in TORCS simulator and use it to collect a dataset of new images with sensors and 5 affordance indicators as compared to 13 indicators in [1]. We then use the dataset that I have generated to develop a CNN (convolutional neural network) algorithm in AlexNet for regression training and self-driving testing. I also trained a GoogLeNet (a different CNN) algorithm and compare it with AlexNet. The training loss of both algorithms converges satisfactorily and testing results show that the self-driving CNN car can successfully run on different tracks unseen by our pre-trained models.

Contents
Abstract i
Acknowledgment iii
1 Introduction 1
1.1 Three General Approaches for Autonomous Driving . . . . . . . 1
1.2 Direct Perception . . . . . . . . . . . . . . . . . . .. . . . 2
1.3 DeepDriving Platform . . . . . . . . . . . . . . . . . . . . . 3
2 System Architecture 5
2.1 TORCS . . . . . . . . . . . . . . . . . . .. . . . . . . . . . 5
2.2 CAFFE . . . . . . . . . . . . . . . .. . . . . . . . . . . . . 6
2.3 Autonomous Driving . . . . . . . . . . . . . . . . . . . . . . 6
3 Controllers 8
3.1 Bernhard’s Controller . . . . . . . . . . . . .. . . . . . . . 8
3.2 DeepDriving Controller (13 Indicators) . . . . . . . .. . . .. 9
3.3 Our Controller (5 Indicators) . . . . . . . . . . . .. . . . . 9
4 Parameter Definitions 11
5 Data Generation 14
5.1 Data Collecting Procedure . . . . . . . . . . . . . . . . . . 14
5.2 Key Elements in Data Generation . . . . . . . . . . . . . . . 15
6 CNN Models 17
6.1 AlexNet . . . . . . . . . . . . . . . . . . . . . . . . . . . 17
6.2 GoogLeNet . . . . . . . . . . . . . . . . . . .. . . . . . . .18
7 Results 20
7.1 Training Results . . . . . . . . . . . . . . . .. . . . . . . 20
7.2 Autonomous Driving Performance Evaluation . . . . . . . . . . 22
7.3 Comparison of MAE between AlexNet and GoogLeNet . . . . . . . 24
8 Conclusions 26
Appendix-Controller Algorithms 27
References 30

References

[1] C. Chen, A. Seff, A. Kornhauser, and J. Xiao, DeepDriving: Learning affordance for direct perception in autonomous driving, Proc. IEEE Int. Conf. Comput. Vis., 2722-2730, 2015.

[2] C. Chen, Extracting Cognition out of Images for the Purpose of Autonomous Driving, Ph.D. Thesis, Princeton University, USA, 2016.

[3] M. Al-Qizwini, et al., Deep learning algorithm for autonomous driving using GoogLeNet, IEEE Intelligent Vehicles Symposium (IV), 89-96, 2017.

[4] B. Wymann, et al., TORCS: The open racing car simulator, Software available at http://torcs.sourceforge.net 4.6, 2000.

[5] Wikipedia Caffe (software)
available at https://en.wikipedia.org/wiki/Caffe_(software)

[6] Wikipedia AlexNet
available at https://en.wikipedia.org/wiki/AlexNet

[7] Wikipedia Convolutional neural network
available at https://en.wikipedia.org/wiki/Convolutional_neural_network