Autonomous vehicle technologies have evolved quickly over the last few years, with safety being one of the key requirements for their full deployment. However, ensuring their safety while navigating through highly interactive and complex scenarios remains a critical challenge. To tackle this problem, intention estimation and motion prediction are fundamental. In this work, a method to infer the intentions, based on a Dynamic Bayesian Network (DBN), and predict the motion, using Markov Chains, of the nearby vehicles at unsignalized intersections is proposed. This approach considers all possible corridors of the surrounding traffic participants and takes into account their interactions to infer the probabilities of stopping or crossing the intersection, as well as the probability of being in each of the possible navigable corridors. To achieve this, the DBN is used to model the relationships between the observed states and the unobserved intentions of the nearby agents. The Markov Chain model, obtained from a kinematic model, is used to predict the future motions of the vehicles, taking into account their current state, their inferred intentions, and the uncertainty associated with the prediction. The resulting multi-modal motion predictions are sent to the ego vehicle to navigate through the scene. The proposed method is evaluated in 6 real situations extracted from publicly available datasets and is compared with a model-based and a learn-based baseline models. The results showed that the proposed method outperformed both baselines in terms of accuracy considering the metrics ADE and FDE.

Roundabouts have many benefits when compared with traditional signal-controlled intersections: improve safety, reduce delay, improve traffic flow, are less expensive, and occupy less area. The behavior of traffic participants is full of uncertainties in the real world. An automated system that relies only on its perception is unable to safely enter the roundabout until a large gap occurs or the vehicle approaching has actually left the roundabout or passed the conflict area. In order to improve the driving quality, autonomous vehicles should be able to infer the correct intention at roundabouts as early as possible. In this work, a method to classify the intentions of the surrounding vehicles at unsignalized roundabouts is proposed. For each vehicle at the scene, a Dynamic Bayesian Network is instantiated and the intentions are inferred using a particle filter.

Motion planning is a process of constant negotiation with the rest of the traffic agents and is highly conditioned by their movement prediction. Indeed, an incorrect prediction could cause the motion planning algorithm to adopt overly conservative or reckless behaviors that can eventually become a dangerous driving situation. This paper presents a framework integrating motion planning and interaction-aware motion prediction algorithms, which interact with each other and are able to run in real-time on complex areas such as roundabouts or intersections. The proposed motion prediction strategy generates a multi-modal probabilistic estimation of the future positions and intentions of the surrounding vehicles by taking into account traffic rules, vehicle interaction, road geometry and the reference trajectory of the ego-vehicle; the resulting predictions are fed into a sampling-based maneuver and trajectory planning algorithm that identifies the possible collision points for every generated trajectory candidate and acts accordingly. This framework enables the automated driving system to have a more agile behavior than other strategies that use more simplistic motion prediction models and where the planning stage does not provide feedback. The approach has been successfully evaluated and compared with a state-of-art approach in highly-interactive scenarios generated from public datasets and real-world situations in a software-in-the-loop simulation system.

Before their massive deployment, autonomous vehicles need to prove in complex scenarios such that they can reach human driving proficiency while guaranteeing higher safety levels. One of the most important human traits to negotiating traffic is the ability to predict the future behavior of surrounding vehicles as a basis for agile and safe navigation. This capability is particularly challenging for an autonomous system in highly interactive driving situations, such as intersections or roundabouts. In this paper, a set of techniques to bring a computationally expensive state-of-the-art motion prediction algorithm to real-time execution are presented with the goal of meeting a standard motion-planning algorithm execution frequency of 5 Hz, which is the primary consumer of motion predictions. This is achieved by applying novel and existing parallelization algorithms that take advantage of graphic processing units (GPUs) through the compute unified device architecture (CUDA) programming language and managing to produce an average 5× speedup over raw C++ in the cases studied. The optimizations are then evaluated in public datasets and a real vehicle on a test track.

Dynamic Occupancy Grids are a popular approach for the estimation of the environment in autonomous driving applications. They offer several advantages, such as the capability to precisely address any type of obstacle and the estimation of the free and unknown space. Nevertheless, they lack contextual information and rely on a low-level representation format. Moreover, additional features beyond occupancy and dynamic states are required to achieve a comprehensive situation understanding. This work proposes an Evidential Semantic Lane-grid as a complementary environment representation to tackle these issues. First, the environment description of the Dynamic Occupancy Grid is extended by analyzing various factors that can influence the perception system's ability to estimate the surroundings and lead to model the unknown space. Then, this enhanced environment description is converted into a high-level representation format based on a semantic lane-grid. This format incorporates the context of the road layout and synthesizes the data with semantic labels, which are consistently estimated over time using an evidential filtering method that supports their meaning and prioritization criteria. Experimental results in real-world urban scenarios show the utility and feasibility of the proposed approach.