在真實環境收集駕駛行為資料是相當危險的事。需準備許多預防措施以免在收集資料時發生危險的事。要收集像左右搖晃這種不安全的駕駛行為在現實中更是困難許多。利用模擬的環境來收集資料既安全又方便，但模擬環境與現實仍有一段差距，因此我們無法將在模擬環境建置的模型使用在真實環境中。為了我們的研究，我們調查了Wasserstein Generative Adversarial Network with Gradient Penalty (WGAN-GP)產生可以用於建置真實環境的模型的現實時間序的駕駛行為資料的能力，目的是希望從模擬的駕駛行為數據合成出近似真實環境的駕駛行為數據。過去WGAN-GP已成功用在與真實影像沒什麼差異的高品質影像。在這份研究中，我們比較一系列生成器的架構，以合成出最好的駕駛行為。三種指標用來評估合成出的資料與真實環境應用(像是駕駛員的身分驗證)中的相似度。最後，我們展示並測量WGAN-GP在產生真實環境中正常駕駛行為的資料有多成功，不過，在產生左右搖晃的資料上仍需改進。;Collecting driving behavior data in a real environment is dangerous dan risky. A lot of precautions need to be prepared to prevent dangerous things happen while doing the data collection. Collecting unsafe behavior such as weaving behavior is even more difficult to do in a real environment. Using a simulation environment is safer and convenient, but there is some gap between simulation and real environment. Hence, we cannot use simulation data to build a model for the real environment. For our research, we investigate the ability of Wasserstein Generative Adversarial Network with Gradient Penalty (WGAN-GP) to produce realistic time series smartwatch-based driving behavior data that can be used to build a model for the real environment. The aim is to generate synthetic driving behavior from simulated driving data similar to real driving data. WGAN-GP has been used successfully to generate a good quality image that indistinguishable from a real image. In this work, we compare a range of generator architecture to generate the best synthetic driving behavior. Three evaluation metrics are then used to quantitatively assess how similar synthetic data is for real-world applications such as driver authentication. Finally, we demonstrate and quantitatively measure how successful WGAN-GP on generating realistic normal-driving data but still need some improvement when generating realistic weaving-driving behavior data.
