林亮志

　　我國汽車運輸業長期面臨交通事故風險高、駕駛管理不易等問題，對道路安全與社會資源造成嚴重衝擊。如何善用新興科技提升營運安全，尤其在駕駛行為風險辨識與預測方面建立有效工具，是政府與運輸業者亟欲解決的關鍵議題。本四年期研究計畫「應用人工智慧分析技術探勘高風險路段」，以人工智慧與車載資料為核心，分階段開發駕駛行為分析工具與風險預警模型，並針對國道客運業者之實務需求進行系統驗證與推廣策略評估。

　　前3年期計畫已完成異常事件的定義與辨識模組建構，整合ADAS警示事件、車內外影像、駕駛行為數據、車道幾何與環境因子等，建立駕駛人風險行為樣態資料庫。第四年期計畫的研究重點主要集中在三大面向：（一）整合風險要素建構駕駛行為預測模型；（二）識別高風險路段與匝道之空間分析；（三）將異常事件分析工具轉化為實務可行之管理系統。首先，在駕駛行為預測方面，計畫採用Boosted Regression Tree與SHAP值分析，依據「本車行為」、「環境互動」、「駕駛員特性與動作」三要素構建多變數風險模型。結果顯示，大多數模型之AUC值達0.8以上，具高度準確性；特定風險如車道偏移、變換車道，與駕駛年資不足、距離前車過近、車道曲率大等情境高度相關，揭示異常事件背後的交互機制。

　　本期導入時間序列資料處理技術與深度學習模型，針對逐秒ADAS警示資料進行預測。測試結果指出，當自變數時間窗設為前10秒、應變數為後5秒時，可有效預測ADAS警示事件發生，且能於駕駛出現高風險動作前即時警示，強化業者事前預防管理能力。此外，SHAP值分析顯示，影響警示事件的重要變數包括車速變異度、方向燈使用、天候狀況與過往警示紀錄，顯示駕駛操作行為與外部環境密切交織影響風險結果。

　　在空間分析方面，計畫針對「巨觀路段」與「巨觀趟次」兩種模式進行建模。巨觀路段分析以2公里為單位，運用負二項迴歸模式建立高風險路段識別工具，反映時空變異下之交通風險分布。巨觀趟次則以15公里為分析單元，建構多層次混合效果模型，處理駕駛人異質性對風險的影響，並導入車流量、班表與駕駛資歷等因素。這些模型不僅提供業者實用的行車前風險提醒，也能協助主管機關掌握潛在高風險區段，做為設置警示標誌與工程改善依據。

　　影像辨識技術部分，本期評估影像辨識輕量化策略，針對車內影像偵測建議採間隔幀取樣以降低系統負荷，惟針對高速國道場景之車外影像辨識仍需維持高幀率，以確保事件連續性與準確度。同時，本期計畫亦納入第二家業者之資料進行系統轉移性評估，發現GPS紀錄頻率與DMS攝影角度為影響資料兼容性的關鍵變數。建議業者若欲採用本系統，應考量升級車機硬體與增設車內影像設備，以提升風險分析品質。

　　為促進技術落地，本期計畫彙整業者、設備商與主管機關意見，提出概念驗證（PoC）、服務驗證（PoS）、商業驗證（PoB）三階段推廣架構，以逐步實現AI風險分析系統的商品化。考量本系統未來在市區道路環境的應用可能性，計畫亦初步評估其可行性，研究結果指出雖核心風險建模邏輯具轉移潛力，但影像辨識模組須因應市區道路遮蔽、車道不明與車種多樣等特性進行調整與再訓練。

　　最終，本期開發完成的駕駛風險管理系統，具備異常事件警示儀表板、駕駛風險綜合報表、趟次風險分數計算與個人行為趨勢追蹤等功能，可協助客運業者即時發現高風險駕駛與路段，並進行針對性管理。此系統除可顯著降低資料處理成本，也提供主管機關一套具科學依據之空間風險分析工具。

　　ABSTRACT:

　　Taiwan’s commercial motor carrier sector has long struggled with a high incidence of road crashes, leading to serious casualties and substantial social costs. In response, this four-year research project titled “Applying Artificial Intelligence Techniques to Identify High-Risk Road Segments” aimed to develop an integrated system for analyzing driving behavior and possible risks, using in-vehicle data and AI technology. The project was executed in phases, progressively enhancing behavior analysis tools and validating them in collaboration with intercity bus operators. The focus of the fourth year was to consolidate data modeling, identify spatial risk patterns, optimize prediction tools, and translate analytical findings into practical safety management solutions.

　　In the first three years, the project established a foundation for risk analysis by developing a set of abnormal event definitions, collecting large-scale telematics and video data, and building the initial behavioral risk indicators. The fourth-year research was structured around three major objectives: (1) integrating multidimensional factors to construct predictive models of risky driving behavior, (2) conducting spatial analysis to identify high-risk road segments and ramps, and (3) transforming analytical outputs into an operational management system. Using Boosted Regression Trees and SHAP, the research team modeled risk based on three domains: vehicle behavior, environmental interaction, and driver characteristics and in-cabin actions. The models demonstrated strong predictive performance, with AUC values consistently reaching 0.8 or higher. Specific behaviors—such as lane deviation or unsafe lane changes—were found to be highly correlated with contextual factors such as insufficient experience, short headways, sharp curves, and mixed traffic conditions.

　　To enable proactive intervention, the team implemented time-series modeling using deep learning algorithms (LSTM, RNN, 1D CNN-LSTM) to predict ADAS warning events in real time. With the optimal configuration—predictor windows of 10 seconds and a response window of 5 seconds—the models successfully forecasted warning events, allowing for risk alerts a few seconds before risky behaviors occurred. SHAP analysis further identified influential variables, such as speed variability, use of turn signals, urban context, previous alerts, and weather, thereby informing targeted interventions and driver feedback strategies.

　　On the spatial analytics front, the project employed two complementary models: a “macro trip model” based on 15-kilometer driving segments and a “macro road segment model” using 2-kilometer units. The trip model incorporated multilevel mixed-effects regression to account for driver heterogeneity, while the segment model applied panel data negative binomial regression to capture monthly variations in traffic risk. These models produced cumulative risk scores, enabling fleet operators to issue pre-trip safety warnings and allowing government agencies to identify segments in need of engineering improvement or focused enforcement.

　　In terms of image recognition technology implementation, the team assessed lightweight strategies for video processing. For in-cabin gesture detection (e.g., steering behaviors), frame sampling effectively reduced computational load without sacrificing performance. However, in highway environments requiring detailed motion detection, full-frame processing remained necessary to preserve temporal continuity. To test system portability, the fourth-year project also conducted a transferability analysis using data from a second operator. Results indicated that GPS logging frequency and DMS (Driver Monitoring System) camera angle were key determinants of compatibility. For full system functionality, operators were advised to upgrade hardware and ensure high-resolution, forward-facing driver views.

　　To support future deployment, stakeholder engagement sessions were conducted with bus operators, equipment suppliers, and regulatory agencies. Based on feedback, a three-stage implementation roadmap was proposed—Proof of Concept (PoC), Proof of Service (PoS), and Proof of Business (PoB). The current system, having reached the PoC stage, includes driver-level dashboards, real-time alerts, and behavior trend reports. Field testing in cooperation with operators is expected to follow, with the PoS phase focused on embedding the system into daily fleet operations. The fourth-year project also considered the applicability of the system to urban driving contexts. While the conceptual framework—including event definition and risk scoring—remains valid, technical adaptations are required. Lane detection and spatial grids need refinement to handle occlusion and high-density mixed traffic, particularly involving smaller vehicles such as motorcycles. Therefore, technology transfer to urban settings will require retraining image recognition models and revising detection logic to account for urban roadway complexity.

　　Finally, the project culminated in the development of a functional Driver Risk Management System, equipped with modules for visualizing abnormal events, tracking warning trends, scoring individual driver behavior, and issuing pre-trip alerts. This system helps fleet operators quickly identify high-risk drivers and road segments, significantly reducing the cost and effort required for data processing and analysis. For regulatory bodies, it provides a scientifically grounded platform for spatial risk profiling and targeted safety interventions.