Fortifying visual-inertial odometry: Lightweight defense against laser interference via a shallow CNN and Optimized Kalman Filtering

Abstract

Accurately estimating a vehicle's position and velocity in real-time is crucial for navigation, especially in environments where updates must be continuous and reliable. Achieving high-precision localization often requires integrating multiple positioning sources, particularly in indoor settings where satellite-based navigation is impractical. Visual-Inertial Odometry (VIO) systems are commonly employed for this purpose. With the rapid advancements in artificial intelligence and deep learning, particularly the deployment of deep networks on GPU-equipped platforms, VIO systems have seen widespread adoption in recent years. However, these systems are vulnerable to remote attacks, such as laser-induced disruptions that can impair camera lenses, leading to compromised image capture and degraded visual localization. This paper introduces a highly-robust system utilizing a shallow convolutional neural network and a fully connected detection layer to enhance VIO resilience against such threats. Moreover, for power-sensitive applications, such as those relying on batteries, an Optimized Kalman Filter (OKF) is used to merge two distinct positioning sources, offering a more efficient alternative to recurrent neural networks like LSTMs. The proposed system demonstrates a 13.27% improvement in accuracy over existing robust VIO systems designed to counteract noise and distortion.

Keywords

• VIO
• IMU
• Laser attack
• Image disruption
• Detector neural network
• Shallow CNN