Scalable Multi-Camera VIO with Bounded Sliding-Window Optimization

Autonomous robots are deployed in real-world applications, where reliable localization is a key enabler of safe autonomy. This paper extends Basalt, a stereo Visual-Inertial Odometry (VIO) system, to support multi-camera rigs on embedded edge devices, improving accuracy and robustness in low-texture and dynamic environments. We introduce a multi-camera VIO formulation that leverages Basalt’s sliding-window bundle adjustment to fuse measurements from additional cameras. In addition, we design a TBB-parallelized multi-camera front-end that performs image-pyramid construction and frame-to-frame feature tracking concurrently across synchronized image streams. Overlapping fields of view across cameras improve cross-view consistency and provide stable metric initialization. We evaluate three-, four-, and six-camera configurations on a quadcopter equipped with an NVIDIA Jetson Orin NX in indoor and outdoor scenarios, using motion capture as ground truth for indoor flights. The proposed system achieves end-to-end latency below 30 ms while reducing drift and attaining lower estimation errors than existing methods in most scenarios.