SafePathAI: Uncertainty-Aware Trajectory Prediction for Safe Autonomous Driving

Abstract

Trajectory prediction is safety-critical in autonomous vehicles, UAVs, air traffic control, and maritime navigation, yet existing models are evaluated almost exclusively on displacement accuracy without assessing predictive calibration. This paper presents SafePathAI, a hybrid ensemble framework integrating a Kalman Filter (KF) and an LSTM with temporal attention, fused via confidence-weighted inverse-variance ensemble fusion. Epistemic uncertainty is quantified through three complementary mechanisms: (i) KF covariance propagation, (ii) Monte Carlo (MC) Dropout variance across M=10 stochastic inference passes, and (iii) a novel inter-model disagreement term that amplifies the uncertainty signal when the two models diverge—a reliable indicator of genuinely ambiguous scenarios that neither model alone can detect. A calibrated rejection mechanism abstains when ensemble uncertainty exceeds threshold τ, deferring to human operators. Experiments across circular, figure-eight, and lane-change trajectories show the ensemble achieves the lowest Average Displacement Error (ADE: 0.31–0.35m) with well-calibrated uncertainty: Expected Calibration Error (ECE: 0.043), Negative Log-Likelihood (NLL: −1.82), and rejection coverage of 92.4%, with ensemble uncertainty 4× higher for high-error versus low-error predictions. An ablation study confirms that removing the disagreement term drops coverage below the 90% safety threshold, and a three-way baseline comparison demonstrates consistent improvement at every stage of the framework, establishing SafePathAI as a principled, deployable solution for safety-critical autonomous systems.