| 摘要: | 由於都市化而產生許多的交通問題造成人員傷亡以及經濟損失，於是交通風險預 測與防範已成為重要的議題，隨著物聯網和人工智慧的發展，智慧運輸系統已成為現 今改善交通的趨勢，人們不但可以利用無線通訊設備即時的溝通以取得最新的路況資 訊，也能透過大數據分析預測未來的交通情況。
然而多數研究已利用交通時空資料來實現交通風險預測，但缺乏考慮時空複雜性， 導致無法有效提取交通狀態特徵。有鑑於此，本研究首先利用皮爾森相關係數探討並 證明交通事故與時間和空間具有高度關聯性，再基於 CNN-LSTM 模型採用多特徵之交 通時空資料進行道路交通風險預測，其中考慮了歷史事故、天氣、交通狀態和時間等 特徵，並結合特徵嵌入方法進行 CNN-LSTM 道路交通風險預測模型之設計與訓練，使 得該模型適用於實際交通環境。在實驗模擬上，本研究針對「單一交通事故資料」與
「多特徵交通時空資料」等兩面向與不同預測模型進行效能分析比較，實驗結果表明， 由於本研究方法基於 CNN-LSTM 模型結合有效提取外部因素的特徵嵌入與 Mask 層， 於是 CNN-LSTM 道路交通風險預測模型可快速收斂且有較低的損失值，此模型具有更 好的道路交通風險預測成效;最後，再進一步探討不同事故因素與交通事故嚴重程度 的關係，本研究比較三種分類器:決策樹、隨機森林和 XGBoost，依用路人族群針對 個人交通事故資料進行分類與分析，從中探討用路人對於交通事故因素的相關性。;With urbanization, many traffic problems have resulted in casualties and economic losses. Therefore, traffic risk prediction and prevention have become important issues. Due to the development of the Internet of Things(IoT)and Artificial Intelligence (AI), Intelligent Transportation System(ITS)have become the trend for improving transporta- tion. People can not only use wireless communication devices to communicate in real-time to obtain the latest road condition information but also predict future traffic conditions through big data analysis.
However, many studies have used traffic spatio-temporal data to achieve traffic ac- cident prediction, but have not considered the spatio-temporal complexity, which leads to the inability to effectively extract traffic features. Given this, this study first uses the Pearson correlation coefficient to explore and prove that traffic accidents are highly cor- related with time and space. Next, we use multi-attribute traffic spatio-temporal data to predict traffic risks based on the CNN-LSTM model, which considers historical accidents, weather, traffic condition, time, and so on. And then we combined the feature embedding method to design and train the “CNN-LSTM Road Traffic Risk Prediction Model”, making the model suitable for the actual traffic environment. In terms of experimental simulation, this study analyzes and compares the performance of “single traffic accident data”and“multi-attribute traffic spatio-temporal data ”with different prediction models. The results show that the CNN-LSTM Road Traffic Risk Prediction Model can converge quickly and effectively with a lower loss value, Our model has better road traffic risk prediction performance; Finally, to further explore the relationship between different ac- cident factors and the severity of traffic accidents. This study compares three classifiers: decision tree, random forest, and XGBoost to classify and analyze the personal traffic accident data of drivers, and then explore the importance of traffic accident factors. |
