自駕車需要通過分析周圍的場景，了解到其他的交通工具(例如:其他車輛、摩托車、腳踏車與行人)的行為並預測其未來軌跡，在複雜的交通環境中高效率地導引自駕車。自駕車主要的目標是規劃出安全的動作並減少可能發生危險時的反應時間。然而，為了學習任何用於交通預測上的模型，都非常需要不同交通事件的訓練圖像。在這項研究中，本論文專注於車禍的圖像合成。為了解決該問題，本研究提出一種基於CycleGAN(循環式生成對抗網路)的框架，以增加車禍的圖像數據。基於學習過的CycleGAN模型，給定車禍圖像，我們的模型可以生成同一種車禍事件於不同場合的圖像。本研究已經證明，所提出的方法在生成圖像品質上優於其他知名的生成對抗網路模型，包括CGAN(條件式生成對抗網路)，DCGAN(深度卷積生成對抗網路)和原始CycleGAN。期望所提出的框架可以促進車禍圖像的數據增強。生成的圖像將有助於深度模型訓練，以進一步應用交通預測，例如:車輛軌跡預測和交通事故預測。

Self-driving vehicles need to efficiently and continuously navigate in complex traffic environments by analyzing the surrounding scene, understanding the behavior of other traffic-agents, such as other vehicles, motorcycles, bicycles, and pedestrians, and predicting their future trajectories. The main goal is to plan a safe motion and reduce the reaction time for possibly imminent hazards. However, to learn any models for traffic prediction, training images of different traffic events are highly required. In this study, we focus on the synthesis of images for car accidents. To solve the problem, a CycleGAN (cycle-consistent generative adversarial networks)-based framework is proposed to augment image data of car accidents. Based on the learned CycleGAN model, given a car accident image, our model can generate several images of different scenes with the same car accident event. The proposed method has been shown to outperform the state-of-the-art GAN architectures, including the CGAN (conditional generative adversarial networks), the DCGAN (deep convolutional generative adversarial networks), the original CycleGAN in terms of the synthesized image quality. It is expected that the proposed framework can facilitate the data augmentation of car accident images. The generated images would be useful to deep model training for further applications of traffic predictions, such as vehicle trajectory prediction and car accident prediction.

目錄
摘要 i
ABSTRACT ii
目錄 iii
圖目錄 v
表目錄 vii
第壹章、簡介 1
1.1 動機與目的 1
1.2 文獻回顧 2
1.3論文架構 4
第貳章、相關研究 5
2.1 卷積神經網路(Convolution neural network) 5
2.1.1 卷積層(Convolution layer) 5
2.1.2 池化層 (Pooling layer) 6
2.1.3 全連接層(Fully Connected Layer) 6
2.2 GAN (Generative Adversarial Network) 7
2.2.1 生成器(Generator) 8
2.2.2 判別器(Discriminator) 9
2.2.3 目標函數 9
2.2.4 反卷積 10
2.3 CGAN(Conditional Generative Adversarial Network) 10
2.4 Encoder-Decoder 11
2.5 Pix2Pix 12
2.6 CycleGAN (Cycle Generative Adversarial Network) 12
2.6.1 CycleGAN 架構 14
2.6.2 CycleGAN 損失函數 14
2.7 VGG-16 16
2.8 殘差塊 16
2.9 PatchGAN 17
2.10 GAN-train and GAN-test 17
第參章、研究方法 18
3.1 研究方法架構 18
3.2擷取資料方式 19
3.3訓練模型 19
第肆章、研究結果 22
4.1實驗環境 22
4.2實驗數據集 23
4.3實驗結果 24
4.4實驗結果比較 27
第伍章、結論以及未來展望 28
5.1結論 28
5.2未來展望 29
參考文獻 30

參考文獻
[1] J. Y. Zhu, T. Park, P. Isola, and A. A. Efros, “Unpaired image-to-image translation using cycle-consistent adversarial networks,” Proc. IEEE Int. Conf. Computer Vision, Venice, Italy, Oct., 2017, pp. 2223–2232.
[2]I. Goodfellow, J. Pouget-Abadie, M. Mirza, B. Xu, D. Warde-Farley, S. Ozair, A. Courville, and Y. Bengio, “Generative adversarial nets,” Generative adversarial nets,” Advances in neural information processing systems, Montreal, Canada, Dec., 2014, pp.2672-2680.
[3] M. Mirza and S. Osindero, “Conditional generative adversarial nets,” arXiv preprint arXiv:1411.1784,2014.
[4] P. Isola, J. Y. Zhu, T. Zhou, and A. A. Efros, “Image-to-Image Translation with conditional adversarial networks,” Proc. IEEE Int. Conf. Comput. Vis. Pattern Recognit, HI, USA, Jul., 2017, pp.1125-1134.
[5] Y. LeCun, L. Bottou, Y. Bengio, and P. Haffner, “Gradient-based learning applied to document recognition,” Proc. IEEE Int. Conf, CA, USA, Jun., 1998, pp. 2278–2324.
[6] A. Krizhevsky, I. Sutskever, and G. E. Hinton, “Imagenet classification with deep convolutional neural networks,” Advances in neural information processing systems, NV, USA, Dec., 2012, pp. 1097-1105.
[7] K. Simonyan and A. Zisserman, “Very deep convolutional networks for large-scale image recognition,” arXiv preprint arXiv: 1409.1556,2014.
[8] K. He, X. Zhang, S. Ren, and J. Sun, “Deep residual learning for image recognition,” Proc. IEEE Int. Conf. Comput. Vis. Pattern Recognit, NV, USA, Jun., 2016, pp.770-778.
[9] M. D. Zeiler, D. Krishnan, G. W. Taylor, and R. Fergus, “Deconvolutional networks,” Proc. IEEE Int. Conf. Comput. Vis. Pattern Recognit, CA, USA, Jun., 2010, pp.2528-2535.
[10]V. Badrinarayanan, A. Kendall, and R. Cipolla, "SegNet: A deep convolutional encoder-decoder architecture for image segmentation," IEEE Trans. Pattern Anal. Mach. Intell., vol. 39, no. 12, pp. 2481-2495, Dec. 2017.
[11] U. Demir and G. Unal, “Patch-based image inpainting with generative adversarial networks,” arXiv preprint arXiv: 1803.07422,2018.
[12] K. Shmelkov, C. Schmid, and K. Alahari, “How good is my GAN,” Proceedings of the European conference on computer vision (ECCV), MUC, BRD, Sep., 2018, pp.213-229.
[13] Y. Lu, Y. W. Tai, and C. K. Tang, “Attribute-guided face generation using conditional cyclegan,” Proceedings of the European conference on computer vision (ECCV), MUC, BRD, Sep., 2018, pp.282-297.
[14] B. Chang, Q. Zhang, S. Pan, and L. Meng, “Generating handwritten chinese characters using cyclegan, ” IEEE Winter Conference on Applications of Computer Vision (WACV), NV, USA, Mar.,2018, pp. 199-207.
[15] H. Lipson, M. Kurman, “Driverless: intelligent cars and the road ahead. Mit Press”, 2016.
[16] J. Long, E. Shelhamer, and T. Darrell, “Fully convolutional networks for semantic segmentation, ” Proc. IEEE Int. Conf. Comput. Vis. Pattern Recognit, MA, USA, Jun., 2015, pp.3431-3440.
[17] O. Ronneberger, P. Fischer, and T. Brox, “U-net: Convolutional networks for biomedical image segmentation, ”International Conference on Medical image computing and computer-assisted intervention, MUC, BRD, Oct., 2015, pp.234-241.
[18] 張斐張 （民94）。類神經網路。臺北市：東華。
[19] A. Shaban, , S. Bansal, Z. Liu, , I. I. Essa and B. Boots, “One-shot learning for semantic segmentation,” arXiv preprint arXiv:1709.03410,2017.
[20] 王維嘉（民108）。AI背後的暗知識：機器如何學習、認知與改造我們的未來世界。臺北市：大寫出版。
[21] C. Balakrishna, , S. Dadashzadeh and S. Soltaninejad,” Automatic detection of lumen and media in the IVUS images using U-Net with VGG16 Encoder”, arXiv preprint arXiv:1806.07554, 2018.
[22] A. Almahairi , S. Rajeswar, A. Sordoni, P. Bachman and A. Courville, “Augmented cyclegan: Learning many-to-many mappings from unpaired data”, arXiv preprint arXiv:1802.10151, 2018.
[23]J. Lin, Y. Xia, T. Qin, Z. Chen, , and T. Y. Liu, “Conditional image-to-image translation, ” Proc. IEEE Int. Conf. Comput. Vis. Pattern Recognit, UT, USA, Jun., 2018, pp.5524-5532.
[24] X. Wang, H. Yan, C. Huo, J. Yu, and C. Pant, “Enhancing Pix2Pix for remote sensing image classification, ” Proc. IEEE Int. Conf. Comput. Vis. Pattern Recognit, Beijing, China, Aug.,2018,pp.2332-2336.