Physics-Informed Neural Network-Based Open Set Classification of Neighboring Vehicle Motion for Decision-Making in Autonomous Driving

Abstract

In this paper, a method for inferring the motion intentions of a neighboring vehicle ahead of an ego vehicle using a physics-informed deep neural network-based open-set classification approach is proposed. Relative motion data from real-world driving were categorized into known and unknown scenarios, with key feature trajectories represented as spatiotemporal 3D input data. A convolutional long short-term memory architecture was designed, and a novel loss function was proposed to incorporate physics-informed perspective and constraints, with integrated loss function's convergence demonstrated for training. The proposed method was evaluated against comparative five classifiers in terms: 1) classification accuracy for known classes in single scenarios; 2) open-set classification accuracy; 3) analysis by deep reduced feature visualization; 4) generalization performance to unknown data; and 5) classification robustness and in-path decision validity in continuous scenarios. Results showed a 23.5% average improvement in accuracy, the highest generalization performance, and superior robustness, enabling faster and more reliable in-path detection compared to conventional radar including in ambiguous situations.