| 摘要: | 隨著深度學習電腦視覺技術在自動駕駛系統與機器人視覺的應用，使得執
行速度與判斷準確度要求日益增加，行人路徑預測已經開始成為當前研究焦點，它可以透過預測影像中行人的移動路徑，在自動駕駛安全系統上可以給予系統更多的反應時間，同時也希望可以減少誤判的機率。另外，在機器人應用領域上可以讓機器人預測並理解人類的移動路徑，進而使其與人有更好的互動與協作能力。然而，行人的真實移動具有其物理性質，當行人在行走時，會考慮到周遭環境，例如: 車道、圍牆、草皮、路樹等，同時也會受其他行人影響自身的移動路徑，上述的問題是很難或無法將其量化或數學化。為了有效解決上述問題，本論文會從時間與空間兩個面向來探討與處理，首先時間的部份，透過卷積神經網路(CNN)來學習行人在不同環境的移動特徵，用以預測行人後續的移動路徑，在空間的部份，透過加入環境資訊圖來協助模型能在預測路徑時達成生成合理的預測。簡言之，本論文將結合座標卷積與自動編碼器運用於行人軌跡預測，所提出之方法可以改善預測的準確度，並同時達成實時預測的要求。最後，藉由大量的測試資料來證明所提出方法明顯優於使用時間序列類型之預測模型方法。;With the increasing demand for execution speed and judgment accuracy of deep learning computer vision technology applicated in the automatic driving system and robot vision, the task of pedestrian trajectory prediction has become the research focus bin predicting the moving trajectory of pedestrians by the frame image. The automatic driving safety system can give the system more response time and reduce the probability of prediction error. And in robotics, the robot can understand human movement trajectory so that it can cooperate better. However, the movement of pedestrians is not rigid body motion that exists in the physical properties. When pedestrians walk, they should consider the surrounding environment, such as lanes, walls, lawns, roadside trees, etc., and also interact with other pedestrians. However, those mentioned above cannot be quantified or mathematical, which is currently a critical problem to overcome.
To effectively address the issues mentioned earlier, this thesis will handle them from time and spatial aspects. For the time aspect, we utilize CNN to learn the pedestrian’s trajectory features and predict their future trajectory in different environments. For the spatial aspect, we incorporate scene information maps to assist the model in generating more reasonable results during the trajectory prediction process. In brief, this work proposes a novel method that combines CoordConv with autoencoders for pedestrian trajectory prediction, and it can improve accuracy and generate pedestrian trajectories efficiently, achieving real-time prediction levels. Finally, we demonstrate the feasibility of the proposed method through extensive testing data, and the results are superior to many RNN-based model predictions. |
