Analysis of Driving Behavior in Unprotected Left Turns for Autonomous Vehicles using Ensemble Deep Clustering

The advent of autonomous driving technology offers transformative potential in mitigating traffic congestion and enhancing road safety. A particularly challenging aspect of traffic dynamics is the unprotected left turn-a scenario at an intersection where the vehicle intending to turn left does not have a dedicated traffic signal, posing a risk to traffic safety and efficiency. This study investigates the dynamics of unprotected left turns by employing data-driven techniques that analyze multi-vehicle data and trajectory patterns to decode complex interactions and behaviors that occur during this maneuver. Our research targets the subtleties of driver behavior in these situations, employing a novel Ensemble Deep Clustering algorithm that innovatively categorizes driving behaviors based on a combination of learned representations and clustering advancements. The deep clustering component involves an iterative process that refines behavioral categorization, while the ensemble technique enhances the precision of these determinations. Using the INTERACTION Dataset, the proposed model is trained and evaluated to offer a better understanding of the intricate driving behaviors in unprotected left turns at intersections. Through the quantitative analysis and comparison with the baseline, we show the superiority of the algorithm, and the results are also interpretable. This methodology can be utilized to improve the decision-making of autonomous vehicles in such scenarios, thus improving the safety of autonomous vehicles, traffic efficiency, and realizing human-robot interaction between autonomous vehicles and drivers.