Lightweight resilience: Advancing visual-inertial odometry with deep convolutional networks and an intelligent learnable Kalman filter for defense against laser remote attacks

Abstract

A combination of multiple positioning sources is often employed to achieve precise and high-rate localization, especially in indoor applications where satellite-based navigation systems are not feasible. This paper presents a super-robust Visual-Inertial Odometry (VIO) system designed to counter laser-induced attacks on camera lenses, a critical vulnerability in visual localization systems. The proposed system integrates lightweight Convolutional Neural Networks (CNNs) to enhance robustness against such attacks. To address power constraints in applications such as battery-powered systems, the solution employs an Intelligent Learnable Kalman Filter (ILKF) for fusing multiple positioning sources, offering a more efficient alternative to Recurrent Neural Networks (RNNs). Simulation results show that the proposed system achieves a 23.09 % improvement in localization accuracy compared to pure INS and a 4.01 % improvement compared to existing robust VIO systems under attack conditions. Additionally, the system reduces the negative impact of laser attacks by 96 %, making it highly suitable for precise indoor navigation in environments where satellite-based systems are unavailable.