自動駕駛是近來熱門的研究題目。目前已經有許多關於物體檢測和追蹤研究工作，而研究重點預計會轉移到物體運動的預測上。在不同類型的物體中，行人是更難預測的，因為人類可以快速地改變方向和狀態（例如，行走或停止）。在這項研究中，我們建立了一個名為 PedCross 的系統，使用人類的圖像語義信息來預測行人的行為（即穿越馬路或不穿越馬路）。在這個系統中，首先會使用行人的影像來檢測人體骨骼。然後藉由人體骨骼來提取特徵值，並用於模型的訓練。Random Forest和 LSTM這兩種類型的模型，會被用於人行穿越馬路的預測。為了進一步提高整個系統的效率和正確率，我們在系統中加入了Skip Frame、Head Orientation 和 Warning/Dangerous Zones這些元件。 PedCross 不僅使用真實場景中收集到的 ITRI dataset進行測試，還實際部署在自動駕駛巴士上，並且進行道路測試。道路測試的結果顯示，PedCross 達到了工研院提出的所有要求，並且優於工研院開發的基準系統 Free Space。;The topic of autonomous driving has become a popular research subject recently. Many research works have been on object detection and tracking, but it is expected that the focus will eventually be shifted to the prediction of object movements. Among different types of objects, pedestrian movements are more difficult to predict because humans can change their direction and status (e.g., walking or stopping) quickly. In this research, we build a system, called PedCross, which uses human image semantic information to predicts the behavior of pedestrians (i.e., crossing or not crossing). In PedCross, images of pedestrians are first used to detect skeletons. The features in the detected skeletons are then extracted for model training. Two types of models, Random Forest and LSTM, are considered for pedestrian crossing prediction. To further improve the efficiency and accuracy of PedCross, Skip Frame, Head Orientation, and Warning/Dangerous Zones are integrated. PedCross is not only tested with the collected ITRI dataset but also deployed on auto-driving bus for road test. The road test indicates that PedCross achieves all the requirements set forth by ITRI and outperforms Free Space, a baseline system developed by ITRI.
