What If I Cannot Always See: Misdetection-Aware MPC for Autonomous Vehicles

Abstract

Autonomous Vehicles (AVs) rely heavily on perception systems to understand their operational environment. However, prevailing AI-driven perception algorithms are prone to errors that may propagate through downstream motion planning and control systems, potentially causing safety risks. This paper investigates a critical yet under-explored pattern of perception errors: misdetection, where the perception system intermittently loses track of objects in the environment. Through a car-following case study, we demonstrate that misdetection poses unique challenges due to its intermittent, nonlinear effects. We identify a safety-critical detection-rate threshold, below which collision risk increases sharply. This effect is particularly pronounced for conventional controllers that account for measurement noise but neglect misdetections, leading to unstable decision-making and jerky vehicle motion. To address these challenges, we propose a Misdetection-Aware MPC (MA-MPC) framework that accounts for both misdetections and measurement noise in its chance constraints. This framework employs a Perception Error Model (PEM) to model both types of errors and introduces a novel uncertainty propagation mechanism to predict future perceptual and motion uncertainties of the leading vehicle. Simulation results show that MA-MPC greatly improves safety, reducing collision risk by over 66% relative to noise-only control, while maintaining smooth vehicle motion during nominal cruising. Our work bridges the critical gap between perception uncertainty and safe control, offering a unified methodology to enhance AV reliability in driving scenarios with imperfect perception systems.