MIXO: Mixture Of Experts-Based Visual Odometry for Multicamera Autonomous Systems

Abstract

Visual-inertial odometry is a fundamental technology exploited by autonomous vehicles and mobile robots to determine their position in unknown environments. Indeed, many techniques have been proposed for monocular and stereo cameras. State-of-the-art autonomous vehicles, however, are equipped with multiple cameras covering the entire vehicle environment. We present here MIxture of eXperts Odometry (MIXO), a data-driven, machine learning-based technique that loosely combines odometry outputs from multiple cameras to obtain a more accurate and robust global estimate. It stems from the intuition that each camera provides an optimal vantage point in specific driving scenarios. In MIXO, each camera (or expert) is individually processed by a state-of-the-art visual odometry algorithm (e.g., ORB-SLAM2). Then, the odometry estimates are mixed by a gating network, which selects the locally optimal experts in the current operational conditions and weights their contributions accordingly. MIXO is a lightweight module that can be easily implemented on top of any visual odometry algorithm. Experimental results on real-life data from autonomous vehicles show that MIXO achieves more robust and accurate results than any single camera, reducing the absolute rotational and translation error by 38% and 15%, respectively.