Model-Free Sensor Fusion for Redundant Measurements Using Sliding Window Variance

In this paper, a model-free data fusion method for combining redundant sensor data is presented. The objective is to maintain a reliable tool center point pose measurement of a long-reach robotic manipulator using a visual sensor system with multiple cameras. The fusion method is based on weighted averaging. The weight parameter for each variable is computed using the sliding window variance with N latest observations. After each sliding window, the window length N is updated, and simple transition smoothing is included. For experimental validation, two sets of pose trajectory data from redundant visual sensors were obtained: 1) using a camera located near the tip of a long-reach manipulator running a simultaneous localization and mapping (SLAM) algorithm and 2) marker-based tracking with cameras located near the base of the manipulator. For pose tracking, a fiducial marker was attached near the SLAM camera. The proposed methodology was examined using a real-time measurement setup and offline data analysis using the recorded data. The results demonstrate that the proposed system can increase the overall robustness and fault tolerance of the system, which are desired features for future autonomous field robotic machines.