Visual-Inertial Cross Fusion: A Fast and Accurate State Estimation Framework for Micro Flapping Wing Rotors

Abstract

Real-time and drift-free state estimation is essential for the flight control of Micro Aerial Vehicles (MAVs). Due to the vibration caused by the particular flapping motion and the stringent constraints of scale, weight, and power, state estimation divergence actually becomes an open challenge for flapping wing platforms’ longterm stable flight. Unlike conventional MAVs, the direct adoption of mature state estimation strategies, such as inertial or vision-based methods, has difficulty obtaining satisfactory sensing performance on flapping wing platforms. Inertial sensors offer high sampling frequency but suffer from flapping-introduced oscillation and drift. External visual sensors, such as motion capture systems, can provide accurate feedback but come with a relatively low sampling rate and severe delay. This work proposes a novel state estimation framework to combine the merits from both to address such key sensing challenges of a special flapping wing platform—micro flapping wing rotors (FWRs). In particular, a cross-fusion scheme, which integrates two alternately updated Extended Kalman Filters based on a convex combination, is proposed to tightly fuse both onboard inertial and external visual information. Such a design leverages both the high sampling rate of the inertial feedback and the accuracy of the external vision-based feedback. To address the sensing delay of the visual feedback, a ring buffer is designed to cache historical states for online drift compensation. Experimental validations have been conducted on two sophisticated microFWRs with different actuation and control principles. Both of them show realtime and drift-free state estimation.

3D Mapping for a Large Crane Using Rotating 2D-Lidar and IMU Attached to the Crane Boom

“…IMU has commonly been used for estimating the pose of a sensor in motion. However, under severe vibration, gyroscope and accelerometer measurements exhibit unmanageable sensor drift caused by sensor bias and noise uncertainty [27]. Many lidar-IMU mapping methods, such as LOAM [19], [20] and its extensions [16], [18], [25], [31], employ the Kalman filter to cope with noise in IMU measurements.…”

“…However, there is a high chance that Kalman filter-based linear state estimation will diverge under high vibration, as in the case of crane application. The complementary filter [26] is used for orientation estimation using an IMU and exhibits better estimation accuracy and robustness than the Kalman filter under high vibration [27], [28], [29], [30]. Combined with the structural information of a crane, the complementary filter can be used for sensor pose estimation.…”

“…Considering the challenges mentioned above, we use a slowly-rotating 2D lidar and an IMU as base components and attach them to the crane boom. As the crane boom moves arbitrarily during operations, we estimate the lidar pose for each scan using the IMU using a complementary filter [26] with moving average filtering, which is more robust to severe vibration than Kalman filter-based methods [27], [28], [29], [30]. To further improve the map accuracy, we develop a new pose graph optimization method using planar environmental constraints that naturally exist in construction sites.…”