如今，汽車駕駛安全的蓬勃發展創造了巨大的國內和全球機遇，尤其是先進智慧駕駛輔助系統（ADAS），這是市場上的主要解決方案。其中一個重要部分是基於視覺的輔助系統。在這篇碩士論文中，我將專注於基於視覺的解決方案，使用機器學習方法來預測汽車的未來，即汽車的運動和位置。還將包括對汽車的驅動控制器優化。

驅動任務的解決方案是在模擬器上，該模擬器在給定一組圖像的情況下簡單但可靠，以計算結果作為關鍵駕駛信息預測的重要指標。透過機器學習過程和使用深度卷積神經網絡，這些模型可用於涵蓋各種軌道和圖像。無論白天和黑夜的時間為何，多樣化的情景預測都能很好地發揮作用。

我收集了超過70萬張圖像來測試有關14個指標和5個指標的訓練結果之表現。此過程涵蓋了TORCS AI駕駛汽車其攝影鏡頭生成的50萬張日間圖像和30萬張夜間圖像。

Nowadays, the prosperous development in car driving safety creates huge domestic and global opportunities, especially for Advanced Driver Assistance Systems (ADAS) which is the major solution in the market. And one of the essential parts is vision-based assistance system. In this master thesis, I will focus on the vision-based solution using machine learning methods to predict the future of a car, the motion and the position of a car. The driving controller optimization to the car will also be included.

The solution of the driving task is on a simulator which is simple and reliably produces a set of images for machine algorithms to learn important indicators in autonomous driving. Through the machine learning process and using a deep Convolutional Neural Network, the algorithms can learn a wide range of tracks and images. The CNN algorithms work well in diverse scenarios in day and night time.

I collected more than 700 thousand images to test the performance of pretrained CNN with 14 indicators and 5 indicators. In total, we have 500 thousand daytime images and 300 thousand night images generated by TORCS AI driver car from its camera.

Contents

Abstract -- p.2
(中文)摘要 -- p.3

Acknowledgements -- p.4

I. Introduction -- p.7 ~ p.13
A. Background
B. Software and Application
C. Mapping from an Image to Key Information
D. Purpose and Method

II. Theory -- p.14 ~ p.35
A. Machine Learning
B. Neural Networks
C. CNN (Convolutional Neural Networks) Loss Layer “Euclidean loss”
□ Sorting
D. Shared Memory
E. Driving Controller
F. More Details on Shared Memory
G. Reinforcement Learning

III. Results -- p.36 ~ p.40
A. Training Loss Results
B. MAE Results
□ Average Static MAE for night condition

IV. Conclusion -- p.41
References -- p.42

References

[1] http://torcs.sourceforge.net/

[2] https://caffe.berkeleyvision.org/

[3] C. Chen, A. Seff, A. L. Kornhauser, and J. Xiao. Deepdriving: Learning affordance for direct perception in autonomous driving. Computer Research Repository (CoRR) in arXiv, 2015.

[4] H. H. Huang, T. P. Wang. The study of learning overtaking and blocking behaviors in a simulated car racing games. Master thesis, 2013.

[5] M. Al-Qizwini, I. Barjasteh, H. Al-Qassab and H. Radha. Deep learning algorithm for autonomous driving using GoogLeNet. Intelligent Vehicles (IV) in IEEE, 2017.

Other useful materials

[6] Build Your Torcs Track in 20 Minutes - https://zh.scribd.com/document/6639748/Build-Your-Torcs-Track-in-20-Minutes