A discrete choice latent class method for capturing unobserved heterogeneity in cyclist crossing behaviour at crosswalks
提出一种离散选择潜类模型，用于捕捉自行车骑行者过环岛人行横道时的多层次未观测异质性，基于无人机轨迹数据识别出两类骑行者（快速通过者和谨慎跟随者），模型拟合优于传统方法。
Conflicts between cyclists and motorized vehicles at crosswalks often lead to severe collisions. The varied behaviour of cyclists at these crossings introduces unobserved heterogeneity. Despite this, there is a notable research gap in studying the cyclist behaviour at roundabout crosswalks. To address this gap, we propose a discrete choice latent class method to capture the multi-level latent heterogeneity in cyclists' crossing behaviour at roundabout crosswalks. Latent heterogeneity can be captured at multiple levels: site-level, interaction-level, choice-attribute level, and individual-level. This method, rooted in behavioural theory, aims to provide a deeper understanding of cyclists' crossing decisions, enhancing safety measures at these intersections. We present an application of the proposed method to two publicly available drone datasets of naturalistic road user trajectories at roundabouts, including 8 roundabout sites that exhibit some level of similarity to minimize site heterogeneity. We capture the latent heterogeneity in the cyclists' membership to a distinct behavioural class at two levels using these datasets: the individual level, represented by the speed of the cyclist as they enter the crosswalk, and the interaction level, defined by the presence of vehicles approaching the cyclist. Our findings align with previous studies that emphasize the significance of the initial speed variable in influencing cyclists' subsequent behaviour and decisions. We identified two distinct classes of cyclists. We hypothesize that Class 1 cyclists, whom we refer to as passers, tend to bypass or overtake other road users at the crosswalk, especially in the absence of vehicles, prioritizing speed and efficiency. We also hypothesize that Class 2 cyclists, referred to as followers, exhibit more cautious behaviour, preferring to maintain a steady pace and avoid overtaking, particularly when vehicles are present. The proposed latent class model effectively captures this behavioural distinction, offering a more granular view of cyclists' decision-making processes at roundabout crosswalks. A key finding is that the discrete choice model with a latent class structure outperforms the basic model without it, despite having more degrees of freedom, as it achieves a lower BIC and AIC but improved model fit statistic. This demonstrates that latent heterogeneity can be effectively captured, leading to improved predictions and outperforming the basic non-latent class model. Classifying cyclists into distinct behavioural classes not only enhances cyclist safety at crosswalks but also provides valuable insights for the development of autonomous vehicle-cyclist interactions.