基於電腦視覺科技執法違規影像之交通衝突分析評估行人穿越行人穿越道之安全性

Abstract

行人安全已成為全球都市道路交通安全的重要議題，聯合國《2021–2030 全球道路安全行動計畫》延續 Vision Zero 與 Safe System 的理念，強調應透過系統性措施降低道路交通死亡與傷害。在此趨勢下，各國逐步強化車輛禮讓行人規範，以提升行人之道路安全保護，其中臺灣實施固定距離車輛讓行執法標準，實務上常以車輛前懸進入行人穿越道，且行人位於其行進方向約3公尺範圍內，作為違規審查之重要依據。雖然此固定距離門檻具有明確、可觀察且便於執法之優點，但其是否足以充分保障行人穿越行人穿越道過程之安全，仍有進一步檢驗之必要。 本研究以臺灣臺北市一處號誌化路口自動化交通執法系統所擷取之影片為資料來源，建立以電腦視覺與軌跡分析為基礎之行人穿越道安全評估架構，探討3公尺讓行標準與行人與車輛衝突風險之關係。共分析 2023 年 9 月至 2025 年 3 月間蒐集之 127 段影片，研究對象為右轉車輛與行人於行人穿越道上之互動情形。另外於方法面，本研究結合相機校正、Homography 鳥瞰平面轉換、YOLOv11x 物件偵測、BoT-SORT 多目標追蹤，以及 Kalman filtering 與 Rauch–Tung–Striebel smoothing 軌跡平滑處理，以重建行人與車輛於地面平面座標中的移動軌跡。完成軌跡重建後，進一步透過車輛幾何重建、行人–車輛配對、時空重疊驗證，以及 Time-to-Collision（TTC）、Post-Encroachment Time（PET）與最小分離距離等替代安全指標，評估行人–車輛衝突風險。需注意的是，本研究所使用之違規與非違規標籤係依據警方審查結果，而非研究者對每一事件進行獨立之精確 3 公尺幾何量測。因此，本研究重點在於比較警方執法審查結果與軌跡式風險指標之關係，而非直接判定個案是否精確符合 3 公尺法定距離。 分析結果指出，在 33 件經 TTC 初步篩選之候選衝突中，經軌跡與幾何驗證後共有 32 件被確認為行人–車輛衝突。其中，25 件屬於 Pedestrian-to-Vehicle 類型，7 件屬於 Vehicle-to-Pedestrian 類型，顯示多數安全相關互動發生於行人已進入穿越空間後，車輛仍持續接近或侵入其行進區域。整體分析事件中，車輛前緣與行人之間沿行人行進方向量測之平均距離為 2.87 公尺，接近臺灣3公尺取締認定基準；非違規事件平均為 3.64 公尺，違規事件平均為 2.52公尺，顯示警方審查結果通常與較短之行人與車輛之空間距離有關。 然而，警方審查結果與軌跡式衝突風險並非完全一致。在確認衝突事件中有15.63%，雖具有 TTC ≤ 3 秒之風險特徵，且其軌跡量測距離介於 3.38 公尺至 4.22 公尺，仍未被警方審查為違規事件；相對地，另6.25%雖軌跡量測距離超過3公尺，仍被警方審查為違規事件。此結果顯示，警方審查結果不應被直接解讀為精確3公尺幾何量測，而可能同時受到車輛侵入程度、行人優先狀態、互動時序與整體危險感知之影響。此外，熱點圖分析顯示，部分高風險衝突集中於行人穿越道起步側約 1 公尺範圍內。另本研究亦觀察到「行人先行、車輛緊跟」之互動型態，並形成 PET ≤ 1 秒之高風險情境。 綜上，本研究顯示交通執法影片除可作為違規審查依據外，亦可透過電腦視覺、鳥瞰軌跡重建與替代安全指標，轉化為主動式行人安全診斷資料。本研究指出，臺灣3公尺標準仍具有作為基本執法門檻之實務價值，但在評估實際衝突風險時，宜輔以軌跡式風險指標與互動時序分析。尤其是行人穿越道起步側約 1 公尺範圍內之高風險互動，以及「行人先行、車輛緊跟」所形成之低時間間隔互動情境，亦應納入行人安全評估與執法檢討之考量。另一方面，3 公尺以外至 4 至 5 公尺範圍內之安全相關互動，亦值得作為未來風險緩衝、執法精進與法規檢討之參考。

Pedestrian safety has become a central issue in urban traffic safety management worldwide. Consistent with Vision Zero, the Safe System approach, and the United Nations Global Plan for the Decade of Action for Road Safety 2021–2030, many countries have strengthened policies and enforcement measures to protect vulnerable road users. Within this international road-safety context, pedestrian-yielding rules are intended to provide drivers with clear responsibilities and pedestrians with greater protection at crossings. In Taiwan, pedestrian-yielding enforcement has also been reinforced in recent years. A commonly applied enforcement criterion considers whether the front overhang of a vehicle enters a crosswalk while a pedestrian is located within approximately three meters along the pedestrian’s walking direction. Although this fixed-distance rule provides a clear, observable, and practical basis for enforcement, it remains uncertain whether such a static threshold can fully reflect the dynamic conflict risk between pedestrians and vehicles. This study develops a computer vision- and trajectory-based framework to evaluate pedestrian crosswalk safety using traffic enforcement videos collected at a signalized intersection in Taipei. A total of 127 video clips recorded between September 2023 and March 2025 were analyzed, focusing on interactions between right-turning vehicles and pedestrians. The framework integrates camera calibration, homography-based bird’s-eye-view (BEV) transformation, YOLOv11x object detection, BoT-SORT tracking, and trajectory smoothing to reconstruct pedestrian and vehicle movements. Conflict risk was assessed through vehicle geometry, temporal–spatial overlap verification, and surrogate safety indicators, including TTC, PET, and minimum separation distance. It should be noted that the violation and non-violation labels used in this study were based on police-reviewed enforcement outcomes rather than independent geometric measurements of the exact three-meter distance. Therefore, the analysis focuses on the consistency between enforcement outcomes and trajectory-based risk indicators, rather than directly validating individual legal-distance determinations. The results show that among the 33 TTC-screened candidate conflicts, 32 were verified as pedestrian–vehicle conflicts after trajectory- and geometry-based validation. Of these verified conflicts, 25 were classified as Pedestrian-to-Vehicle conflicts and 7 as Vehicle-to-Pedestrian conflicts. This indicates that most safety-relevant interactions occurred when vehicles continued to approach or encroach upon the pedestrian movement space after pedestrians had already entered the crossing area. Across all analyzed events, the mean vehicle-front-to-pedestrian distance measured along the pedestrian’s walking direction was 2.87 m, which was close to Taiwan’s three-meter enforcement criterion. The mean distance was 3.64 m for police-reviewed non-violation events and 2.52 m for police-reviewed violation events, suggesting that police-reviewed enforcement outcomes were generally associated with shorter pedestrian–vehicle spatial separation. However, police-reviewed enforcement outcomes and trajectory-based conflict risk were not completely consistent. Among the verified conflicts, 15.63% showed TTC values of 3 s or less and had trajectory-derived distances ranging from 3.38 m to 4.22 m, but were not classified as violation events by police review. Conversely, 6.25% were classified as violation events even though their trajectory-derived distances exceeded 3 m. These findings suggest that police-reviewed enforcement outcomes should not be interpreted as precise geometric measurements of the three-meter threshold. Instead, such outcomes may also reflect broader interaction contexts, including vehicle encroachment, pedestrian priority, movement sequence, and perceived overall risk. In addition, hotspot analysis showed that some high-risk conflicts were concentrated within approximately one meter of the near-side crosswalk entry. This study also observed Pedestrian-first, Vehicle-follows interaction patterns, which produced high-risk situations with PET values of 1 s or less. Overall, this study demonstrates that traffic enforcement videos can serve not only as evidence for legal violation review but also as valuable data for proactive pedestrian safety diagnosis when combined with computer vision, BEV trajectory reconstruction, and surrogate safety indicators. The findings suggest that Taiwan’s three-meter criterion remains practically useful as a basic enforcement threshold. However, when evaluating actual conflict risk, it should be supplemented by trajectory-based risk indicators and movement-sequence analysis. In particular, near-side pedestrian exposure within approximately one meter of the crosswalk entry, as well as low-time-gap interaction patterns formed by Pedestrian-first, Vehicle-follows situations, should also be considered in pedestrian safety evaluation and enforcement review. In addition, safety-relevant interactions occurring beyond three meters, particularly within the four- to five-meter range, may provide useful evidence for future risk-buffer evaluation, enforcement improvement, and regulatory review.