Navigating Unsignalized Intersections: A Predictive Approach for Safe and Cautious Autonomous Driving

Collision avoidance at unsignalized intersections is critical to autonomous vehicle technology. Our work addresses the challenging problem of online speed planning along a predefined traverse path based on an interaction-aware prediction algorithm between autonomous and human-driven vehicles in unsignalized intersections. The algorithm is executed in three sequential steps at each instant: (i) An LSTM and a graph neural network (GNN) are trained to predict the sequence of vehicles crossing the collision point based on their interactions, acting as an expert driver. Our future behavior predictor achieved an accuracy of 90.5%, outperforming the alternative approach, which achieved 88%. (ii) Another GNN estimates the approximate time window required for the self-driving car to traverse the collision point safely, based on the sequence of passage obtained in the previous stage. (iii) The algorithm calculates an acceleration value that not only ensures crossing the collision point within the time window but also avoids dangerous situations in subsequent collision points. By considering both the immediate and subsequent collision points, the algorithm promotes a more comprehensive and cautious driving behavior in autonomous vehicles. We ran our algorithm on 250 real-world scenarios from the INTERACTION dataset, and the self-driving vehicle safely passed the intersection in all cases, even more cautiously than human-driven cars. We presented three case studies to illustrate the algorithm's effectiveness. The algorithm is executed every 100 ms, and the average computational time is 28 ms.