在公路上基於深度學習的軌跡預測與雷達模擬

在本文中，我們將針對自駕車這個主題進行研究，使用comma.ai 開源的自駕車系統openpilot 作為我們的研究平台，我們根據openpilot 的模型以及資料集進行研究，透過他們開源的GitHub，我們可以做出一個相似於 openpilot 的模型，透過這個模型，我們可以在未來建構自己的自駕車模型，並在這套成熟的自駕車系統上運行，值得注意的是，我們只考慮在公路上、不換道的簡單且單調的路況進行學習，這個模型主要用於預測未來軌跡以及雷達的資訊。此外，我們比較了原模型(supercombo) 與我們的模型，結果顯示我們的預測結果比supercombo 更準確。我們還提出在模型中使用Attention-based Convolution Neural Network (ACNN)，我們的研究顯示，使用ACNN 架構的軌跡預測效果比openpilot 的遞迴神經網路(GRU) 更好。最後，我們成功地在openpilot 軟體以及硬體平台上執行我們的神經網路。

關鍵字

軌跡預測；雷達模擬；自動駕駛；深度學習

並列摘要

In this article, we will conduct research on the subject of self-driving cars, using comma.ai’s open-source self-driving car system openpilot as our research platform. We conduct research based on openpilot’s model and data set. Through their open source GitHub, we can create a model similar to openpilot. Through this model, we can construct our own self-driving model for predicting car future trajectory and run the model on openpilot installed in a comma.ai device called comma two. Significantly , we only consider simple and monotonous road conditions on the highway without changing lanes. This model is mainly used to predict future trajectories and radar information. In addition, we compared the original model (supercombo) with our model, and the results show that our prediction results are more accurate than supercombo.We also propose to use an attention-based convolutional neural network (ACNN) in the model. Our results show that ACNN is better than the recurrent neural network of openpilot in terms of trajectory prediction. Finally, we successfully excute our neural network on the openpilot software and hardware platform.