Naturalistic Driver Intention and Path Prediction using Recurrent Neural Networks (1807.09995v1)

Abstract: Understanding the intentions of drivers at intersections is a critical component for autonomous vehicles. Urban intersections that do not have traffic signals are a common epicentre of highly variable vehicle movement and interactions. We present a method for predicting driver intent at urban intersections through multi-modal trajectory prediction with uncertainty. Our method is based on recurrent neural networks combined with a mixture density network output layer. To consolidate the multi-modal nature of the output probability distribution, we introduce a clustering algorithm that extracts the set of possible paths that exist in the prediction output, and ranks them according to likelihood. To verify the method's performance and generalizability, we present a real-world dataset that consists of over 23,000 vehicles traversing five different intersections, collected using a vehicle mounted Lidar based tracking system. An array of metrics is used to demonstrate the performance of the model against several baselines.

Naturalistic Driver Intention and Path Prediction Using Recurrent Neural Networks

The paper "Naturalistic Driver Intention and Path Prediction using Recurrent Neural Networks" offers a significant contribution to the field of autonomous vehicle navigation and Advanced Driver Assistance Systems (ADAS). It addresses the challenge of predicting driver intentions at unsignalized urban intersections—a scenario of particular complexity due to its lack of structured road signals and high variability in vehicle movement.

The methodology is centered around recurrent neural networks (RNNs), specifically utilizing a Mixture Density Network (MDN) in conjunction with RNN to capture the inherent uncertainty and multimodal nature of driver behavior. This approach diverges from traditional single-model trajectory predictors, acknowledging that at intersections, multiple possible paths may exist with varying likelihoods. Furthermore, the paper introduces an innovative clustering algorithm, termed Multi-PAC, which extracts the relevant paths from the probability distribution and ranks them according to their likelihood.

Empirical analysis within this paper leverages a substantial dataset collected from over 23,000 real-world vehicle trajectories across five distinct roundabout intersections in Sydney, Australia. This dataset is notably one of the largest of its kind, contributing valuable insights and benchmarks to the field. The performance of the proposed model is examined using several key metrics, including Euclidean and Modified Hausdorff distances—both critical for evaluating the accuracy of path predictions in terms of spatial and temporal alignment with reality.

Several noteworthy results are highlighted:

• The RNN model incorporating the MDN outperforms baseline models such as constant velocity (CV), constant yaw rate and velocity (CTRV), and constant acceleration (CTRA) in predicting vehicular trajectories.
• For left and right-turning vehicles, the RNN-FF model (random feedforward sampling within the predictions) demonstrates superior accuracy and alignment with the actual path compared to historical and Gaussian Process (GP) methodologies.
• Even for complex intersections with less conducive visibility, the model provides multiple hypotheses that encompass possible trajectories, distinctly improving upon single-path prediction systems.

By utilizing real-world data and incorporating uncertainties through probabilistic outputs, the authors provide an application-oriented framework with practical significance in enhancing the safety and operational efficiency of autonomous vehicles in urban environments. This approach is shown to generalize effectively across different intersection types, suggesting its potential utility in a wide array of driving conditions.

While the research does not explore driver interactions with other vehicles extensively, the groundwork presented for path prediction is robust and lays the foundation for future investigations perhaps incorporating social pooling layers or interaction-based augmentation pathways. Such developments could propel this predictive framework towards more comprehensive ADAS and autonomous driving systems capable of navigating the complexities of urban traffic with higher precision.

In conclusion, this paper's contributions reinforce the merits of deep learning and probabilistic modeling in the trajectory prediction field, particularly under conditions rich with uncertainty and variability. The findings underscore a pivotal step toward more adept autonomous systems that adapt dynamically to both structured and unstructured real-world settings. Future research may explore interaction-aware models, potentially informed by the successful methodology delineated in this paper.