結合CNN與LSTM深度學習方法，於行車紀錄器預測車輛碰撞風險之預測

A CNN-LSTM Based Model for Traffic Accident Prediction by Using Dashboard Camera

交通事故件數持續增長，如何預防事故為一重要課題。隨著車輛技術的發展，自動駕駛車輛可藉由感測器（如攝像頭、雷達等）收集大量即時交通資訊，行車紀錄器也普遍應用於釐清肇事責任的工具，即時交通狀況亦可作為事故預測的資料。近年來深度學習已被廣泛作為分類、檢測和辨識的方法。深度學習方法透過深度非線性網路結構得出複雜的近似函數，並使用堆疊的隱藏層自行擷取出特徵以提升準確率。本研究採用深度學習方法，以行車紀錄器事故影片，結合預訓練的CNN 和LSTM 建構預測模型，用以預測車輛在3 秒內車輛碰撞風險。本研究採用卷積神經網路 （convolution neural network, CNN）與長短期記憶 （long short-term memory, LSTM） 分別能有效擷取資料的空間及時間特徵。

More and more traffic accidents have caused huge social costs in recent years. Traffic managers are devoted to improving road safety and preventing traffic accidents. With the advancement of technology, autonomous vehicles can collect a large amount of real-time traffic data through the sensors (such as cameras, radars, etc.). Dashboard cameras have also been commonly installed on vehicles to clarify the responsibility for accidents, and real-time traffic information can be used as general data for accident prediction. Deep learning has been widely used as a classification, detection, and recognition method. Deep learning models improve classification accuracy by using deep nonlinear network structures to realize complex function approximation and extract features spontaneously through multiple stacked hidden layers. This study focuses on the applications of deep learning models in traffic accident prediction. The vehicle collision video recorded by the dashboard camera is collected. A prediction model based on a pre-trained convolution neural network (CNN) and Long Short-Term Memory (LSTM) is developed for predicting vehicle collision risk 3 seconds before the collision. To provide accurate prediction results, the deep learning process in this study includes CNN and LSTM, which are used to extract spatial features and temporal features, respectively.