應用肇事影像探索行車空間位置與肇事風險關聯之研究

A study on the Correlation between the Locations of Vehicles and Accident Events Using Images of the Accidents

温谷琳(Ku-Lin Wen)

指導教授 ： 許添本

國立臺灣大學/工學院/土木工程學研究所/博士(2021年)

https://doi.org/10.6342/NTU202104404

摘要

在台灣，機車涉入的肇事為最主要的肇事問題類型。在交通安全改善，肇事之因子為重要的研究議題。過去，多數研究主要關注於易肇事地點的靜態肇事因子。然而，肇事的動態空間特性卻常被忽略。主要是因為資料取得不易，且很少有關於空間因子與肇事事件關連性的相關研究。本研究旨在探索在交叉口之機車與汽車碰撞，其涉及的空間肇事因子以及其對衝突的影響程度。本研究以路口肇事影像來建立空間肇事因子，並整合肇事事件的影像資料以發掘重要肇事因子與建立肇事風險模式。本研究應用影像資料來分析路口交通車流與肇事，並透過隨機森林與類神將網絡模式，分別找出重要肇事因子。再從兩者的重要肇事因子的排序，以成對樣本排序檢定比較篩選出的重要因子排序是否一致。另外，分別以應用類神經網絡與多項邏輯迴歸分別建立預測模式。由影像分析結果顯示，部分肇事現場圖所記錄的肇事位置，與實際發生的位置並不一致。本研究發現當機動車駕駛選擇斷面不同位置穿越交叉口時，其肇事風險等級不同。從成對樣本排序檢定比較分析的結果，由隨機森林與類神經網絡而得之重要肇事因子排序兩者存在一致性。此外，結合非事件資料而建立的肇事預測模式顯示，當模式包含空間因子時可以增加模式的預測能力至 79.4%。此外，多元邏輯斯迴歸也應用於探索空間因子對於肇事分析的影響。結果顯示，可提供更準確的預測模式。此外，機車在停止線、機車在衝突點、左轉汽車在衝突點的穿越區位，顯著的影響肇事風險模式。
本研究發現傳統分析方法，在某些肇事中無法發現實際肇事原因。透過影像分析，碰撞位置可以被精確的確認。由車流分析顯示，左轉汽車在斷面的分佈空間位置，是複雜且不規則的。因此，要考慮左轉車流特性對於肇事風險之影響是困難的。雖然過去的研究所探索的肇事因子可以作為交通單位的參考，但是過去的研究無法提供在衝突過程中，機動車駕駛的軌跡行為。本研究發現，透過確認與分析肇事影像，可以瞭解機動車穿越交叉口的位置顯著的影響肇事風險。當直行機車選擇行駛於右側，則容易與右轉汽車衝突，進而導致直行機車行駛的軌跡與速度產生變化。若左轉汽車未適當地停等，並以適合的轉向軌跡轉向，左轉汽車便無法在預期的衝突區穿越交叉口。此情況增加的機車與對向左轉汽車的碰撞風險。本研究提供了交叉口重要的設計資訊以改善機車肇事。此外，對於自駕車而言，本研究也可應用於自駕車未來在混合車流環境中，針對防撞相關參數設計的參考。

關鍵字

資料探勘 ； 交通安全 ； 機車 ； 肇事因子 ； 隨機森林 ； 類神經網絡 ； 多元邏輯斯迴歸

並列摘要

In Taiwan, the motorcycle-involved accident is the primary type of motor vehicle accident. A crucial issue in traffic safety enhancement is the factors that influence accidents. In past research, most studies have mainly focused on static accident factors in hot spots. However, the dynamic spatial characteristics of accidents are usually ignored because of the difficulty of obtaining data and a lack of research on the relationships between spatial factors and accident events. The purpose of this research was to explore the spatial factors of intersections where motorcycles collide with cars and their degrees of impact on the crash. In this study, images of accidents at intersections were used to establish spatial accident factors and other incident data were integrated to find important accident factors and build an random forest model and artificial neural network model.
The research applied video recording data to analyze traffic flows and accidents at the intersections. By using the random forest and artificial neural networks model are applied to identify the critical factors responsible for traffic accidents. From the important factors, the factors’ rankings were compared by paired sample ranking tests. Artificial neural networks and multinomial logistic regression model are respectively developed for accident predictions.
The video recording analysis results showed that some accident locations of the crash scene diagram were not the same as the real locations. In this study, it is found that vehicle drivers choosing different locations of the cross-section incur different accident risk levels while crossing the intersection. Paired sample ranking tests indicated consistency in the ranking of the important accident factors from the results of the random forest and artificial neural networks analysis. The accident prediction models with non-incident data showed that including spatial factors increased the accuracy of the model. In addition, multinomial logistic regression was utilized to explore the influences of spatial accident factors on crash risks. The results showed that the model integrating spatial factors provided a more precise prediction outcome. In addition, the accident risk is significantly affected by the location in the cross-section where a motorcycle crosses the stop line, the conflict location, and the corresponding conflict zone for left-turn cars.
This study finds that traditional analysis methods cannot find the real causes of some accidents. By accident video analysis, the collision location can be accurately determined. According to the traffic flow analysis, the distribution of the cross-sectional space of the left-turn cars shows that the trajectory of the left-turn cars is complex and irregular at intersections. Therefore, it is hard to consider the influence of this left-turning traffic flow characteristic in the accident risk. Although the accident factors from past research can serve as a reference for traffic authorities, past research does not provide the vehicles’ drivers trajectory behavior during conflicts. This research found, by defining and analyzing the accident video, that the location of the vehicle when passing through the intersection affects the accident risk significantly. When a straight motorcycle chooses to ride on the right side, then it will easily conflict with right-turn cars, resulting in variations in the riding trajectory and speed of the straight motorcycle. If the left-turn car does not wait properly and turn with an appropriate turning trajectory, the left-turn car cannot turn through a predictable conflict area. Such situations increase the collision risk between motorcycles and opposing left-turn cars. Finally, this research could provide vital information for intersection design to reduce the frequency of motorcycle accidents. In addition, for autonomous vehicles, the findings of this research could be used as a reference for the design of anti-collision parameters in a mixed traffic environment in the future.

並列關鍵字

data mining ； traffic safety ； motorcycle ； accident factor ； random forest ； artificial neural network ； multinomial logistic regression

參考文獻

[1] P. Chen, W. Zeng, G. Yu, and Y. Wang, “Surrogate safety analysis of pedestrian-vehicle conflict at intersections using unmanned aerial vehicle videos,” Journal of advanced transportation, vol. 2017, 2017.

[2] National Accident Database. 2019.

[3] A. Montella, “A comparative analysis of hotspot identification methods.,” Accident; analysis and prevention, vol. 42, no. 2, pp. 571–581, 2010, doi: 10.1016/j.aap.2009.09.025.

[4] Ž. Šarić, G. Zovak, and N. Koronc, “Comparison of methods for determining crash hotspots in the road traffic,” 2011.

[5] 蘇志哲, 林宜達, and 王雲雱, 易肇事地點改善作業技術參考手冊. 交通部運輸研究所, 2003.