社會力場模式建構市區道路環境下機車移動行為模式

隨著自駕車技術的發展，自動駕駛與人為駕駛之差異性會影響道路車流環境，因而產生模擬評估其影響狀況的需求。但現有的成熟模擬平台皆以汽車行為框架來描述機車行為，經文獻回顧探討發現，汽車之車道模型不足以捕捉機車的特殊行為。因此模擬平台所建構的環境與台灣市區道路車流之間存在落差。所以本研究希望建構機車移動行為模式，用以預測機車於下時階位置，並以領地空間概念探索機車於空間中的特性。本研究透過圖像分析與統計方法，確立領地空間的最佳組合，分別為分群角度4度與兩車間距離90%統計值，並較適合以橢圓形簡化該空間。在探索過程中，發現領地空間受本車車速影響，會隨著速度增加而擴張；鄰車會因機車本車不同角度方位劃分區域而有所差異；領地空間也會因駕駛習慣而產生偏差角度，但角度差異並不明顯而可忽略左右側的誤差，上述的結論也作為後續模型建構的前提假設。本研究認為機車移動方式更近似於行人的行為，機車較不受制於車道線束縛而靈活多變。因此延續人流模型中社會力場模型，將其應用於預測機車移動行為。模式中假設機車的周圍會形成力場，場中以加速度向量建構與影響對象間的交互關係，相比於過去類似機車模型，本研究提出淨空間引力概念，認為機車移動過程中會受到道路淨空間所吸引，而非過往所假設的影響車輛，更能解釋機車於自由車流下自主加速的行為。模式參數校估上提出一種結合基因演算法(GA)與基於模擬的推理(sbi)之兩階段架構，在求得概似率最高的參數組合下，進一步得出各參數之後驗分布，有助於理解模式的參數特性，同時能夠捕捉不同駕駛個體間的差異性。本研究以加速度分布的中央趨勢來衡量模型有效性，並計算速度與轉向角之RMSE與sMAPE評估模式預測能力，最終測試結果證明該模式有效且預測能力良好。藉由核密度估計分布圖發現，所提出的兩階段架構校估方法相較於單基因演算法(GA)校估出的參數，更貼近於觀測實際樣本分布。本研究成果未來可透過模擬平台所提供接口自定義機車行為，以建構較符合台灣車流之模擬環境，供後續自駕車相關研究。

關鍵字

領地空間；社會力場模式；機車移動行為；微觀車流模擬；無車道模式

並列摘要

With the development of self-driving car technology, the differences between autonomous driving and human driving can affect the traffic flow environment, creating a need to simulate and assess their impact. However, existing mature simulation platforms describe motorcycle behavior using a car behavior framework, and a literature review has revealed that the lane model for cars is insufficient to capture the unique behavior of motorcycles. As a result, there is a discrepancy between the simulated environment and the actual traffic flow on urban roads in Taiwan. Therefore, this study aims to construct a motorcycle movement behavior model to predict the future positions of motorcycles and explore their characteristics using the concept of territorial space. Through image analysis and statistical methods, this study establishes the optimal combination of territorial space, including a clustering angle of 4 degrees and a statistical value of 90% for inter-vehicle distance. The territorial space is best approximated using an elliptical shape. During the exploration process, it was found that the territorial space is influenced by the motorcycle's own speed and expands with increasing speed. The neighboring vehicles have different regions based on the angular position of the motorcycle, and the territorial space can also deviate due to driving habits. However, the angle differences are not significant enough to consider the errors on the left and right sides, and these conclusions serve as the underlying assumptions for subsequent model construction. This study suggests that motorcycle movement is more similar to pedestrian behavior, as motorcycles are less constrained by lane markings and exhibit greater flexibility. Therefore, building on the social force model used in pedestrian flow models, this study applies it to predict motorcycle movement behavior. The model assumes that a force field forms around the motorcycle, and the interactions between objects are constructed using acceleration vectors. Compared to previous motorcycle models, this study introduces the concept of free space attraction, proposing that motorcycles are attracted by the empty space on the road rather than being influenced by other vehicles. This concept better explains the autonomous acceleration behavior of motorcycles in free-flowing traffic. The model parameter calibration proposes a two-stage framework combining genetic algorithms (GA) and simulation-based inference (sbi). By obtaining the parameter combination with the highest likelihood, the posterior distribution of each parameter can be determined, aiding in understanding the characteristics of the model's parameters and capturing individual differences among different drivers. The effectiveness of the model is evaluated based on the central tendency of the acceleration distribution, and the predictive ability of the model is assessed using the root mean square error (RMSE) and scaled mean absolute percentage error (sMAPE) for velocity and steering angle. The final test results demonstrate the effectiveness and good predictive capability of the model. By examining the kernel density estimation distribution plot, it is found that the two-stage framework calibration method proposed in this study provides parameter estimates that are closer to the observed actual sample distribution compared to parameter estimates obtained through a single genetic algorithm (GA). The results of this study can be used to customize motorcycle behavior in simulation platforms, creating simulation environments that better reflect the traffic flow in Taiwan for future research related to autonomous vehicles.