建構交通事件之人工智慧物件偵測邏輯與實證研究

The Development of AI Object-detection Logics and an Empirical Study for Traffic Incident

道路交通事件資訊整合與發布平台可有效地讓交控中心即時地執行交通管理並讓用路人藉由重新規劃路徑以提升便利性。目前人工智慧技術被廣泛應用在各個領域，但鮮少研究致力於在交通事件偵測領域上之應用。因此，本研究嘗試發展一套人工智慧深度學習物件偵測演算邏輯，包括紅線違停、異常停留、路口溢流、逆向行駛等常見交通事件，用以自動化偵測與辨識交通事件。實證分析結果顯示，異常停留之整體準確率為73.33%、紅線違規停車之整體準確率為95.95%%、路口溢流偵測之整體準確率為99.40%、逆向行駛之整體準確率為39.28%。本研究進而探究人工智慧交通事件偵測準確率不盡理想之結果，研擬精進模式準確率之方法。

The integration and announcement platform of road traffic incidents can allow traffic control centers to effectively implement real-time traffic management and enhance travelers' trips by rerouting them where necessary. Nevertheless, there are not enough studies that can fully explain how artificial intelligence (AI) can be utilized to make a positive difference in this industry. In order to fill this need, this study has devised AI logics of object detection, along with a deep learning neural network, in order to identify and record traffic misdemeanors. This can then enable the traffic department to efficiently control and monitor how motorists behave while driving. Factors that will be monitored in the proposed AI-based traffic incident logics are illegal parking along a red line, parking violations, intersection overflows, and driving in the wrong direction. In order to test the effectiveness of this AI-based method, a field case was conducted with the following results. The empirical results reveal that the overall accuracy of parking violations is 73.33%, the overall accuracy of illegal parking along a red line is 95.95%, the overall accuracy of intersection overflows is 99.40%, and the overall accuracy of driving in the wrong direction is 39.28%. This study further examines which components of the AI traffic incident detection might be inexact and outlines potential approaches that could enhance the accuracy of these models.