DAP: Dynamic Adversarial Patch Attack to Feature-based Visual Odometry

Visual odometry is a fundamental task in autonomous driving, providing vehicle poses that serve as essential inputs to higher-level modules. In this study, we innovatively propose Dynamic Adversarial Patch (DAP) attack targeted at the widely adopted visual odometry algorithms. Unlike previous attacks, our approach deploys specially designed adversarial patch at vulnerable locations within the scene and move the content in a fixed direction, thereby inducing errors in pose estimation. We evaluated the efficacy of the attack simulated environments. The experiment results show that the our attack can cause serious deviation on feature-based visual odometry.