A Decentralized Sensor Fusion Scheme for Multi Sensorial Fault Resilient Pose Estimation

Abstract

This article proposes a novel decentralized two-layered and multi-sensorial based fusion architecture for establishing a novel resilient pose estimation scheme. As it will be presented, the first layer of the fusion architecture considers a set of distributed nodes. All the possible combinations of pose information, appearing from different sensors, are integrated to acquire various possibilities of estimated pose obtained by involving multiple extended Kalman filters. Based on the estimated poses, obtained from the first layer, a Fault Resilient Optimal Information Fusion (FR-OIF) paradigm is introduced in the second layer to provide a trusted pose estimation. The second layer incorporates the output of each node (constructed in the first layer) in a weighted linear combination form, while explicitly accounting for the maximum likelihood fusion criterion. Moreover, in the case of inaccurate measurements, the proposed FR-OIF formulation enables a self resiliency by embedding a built-in fault isolation mechanism. Additionally, the FR-OIF scheme is also able to address accurate localization in the presence of sensor failures or erroneous measurements. To demonstrate the effectiveness of the proposed fusion architecture, extensive experimental studies have been conducted with a micro aerial vehicle, equipped with various onboard pose sensors, such as a 3D lidar, a real-sense camera, an ultra wide band node, and an IMU. The efficiency of the proposed novel framework is extensively evaluated through multiple experimental results, while its superiority is also demonstrated through a comparison with the classical multi-sensorial centralized fusion approach.

Keywords: decentralized fusion; fault resilient optimal information fusion; linear minimum variance; maximum likelihood function; multi sensor fusion; optimal information filter.

Figure 4

Combination of position and orientation information flow from various sensors in constituting different nodes of the first layer.

Figure 1

The aerial robot considered with the heterogeneous and asynchronous sensors for establishing the multi-layer sensor fusion architecture.

Figure 2

Co-ordinate frames: subscript W denotes the global frame and subscript B denotes the body frame.

Figure 3

The Fault resilient optimal information filter with a two-layered fusion arrangement. An emphasized description of the first layer fusion is given in Figure 4.

Figure 5

Variation of estimated MAV trajectory obtained from Centralized Fusion (CF), decentralized Optimal Information Fusion (OIF), Fault Resilient (FR)-OIF, and ground truth (position from vicon camera). FR-OIF provides the estimated position of the MAV approximately close with the ground truth obtained from the Vicon motion capture system.

Figure 6

Estimated positions along ‘X-Y-Z’ components obtained from intermediate nodes ($$A,\cdots ,G$$) and FR-OIF, compared with CF, OIF and the ground truth. The evaluation is carried out in presence of temporal fault appearing from LIO (Case-1). Except the fused position obtained from FR-OIF all the estimated positions deviated during 20–30 s.

Figure 7

Variation of orientation represented using Euler angles obtained from intermediate nodes ($$A,\cdots ,G$$), CF, OIF, FR-OIF and the ground truth. During the experiment, the transnational motion is dominant over the rotational motion. As a result, the significant impact of FR-OIF is difficult to be visualized.

Figure 8

Case-1: Comparison of estimated position obtained from CF, OIF, FR-OIF and Vicon based ground truth, visualization in an emphasized mode of Figure 5 is presented. In the presence of a fault in the LIO for the duration of operation between (20–30) s, the centralized and classical OIF approach is unable to recover the failure in the estimated states, while the proposed FR-OIF successfully recovered from the faulty measurements and it is able to provide a close approximation of position estimate comparable with ground truth.

Figure 9

Case-2: Estimated positions along ‘X-Y-Z’ components obtained from intermediate nodes ($$A,\cdots ,G$$) and FR-OIF, compared with CF, OIF and the ground truth. The evaluation is carried out in presence of temporal fault appearing from VIO. Except the fused position obtained from FR-OIF all the estimated positions deviated during 50–60 s.

Figure 10

Case-3: Estimated positions along ‘X-Y-Z’ components obtained from intermediate nodes ($$A,\cdots ,G$$) and FR-OIF, compared with CF, OIF and the ground truth. The evaluation is carried out in presence of temporal fault appearing from UWB (20–30) s and LIO (35–45) s. Except the fused position obtained from FR-OIF all the estimated positions deviated in presence of faulty measurements.

Figure 11

Case-4: Estimated positions along ‘X-Y-Z’ components obtained from intermediate nodes ($$A,\cdots ,G$$) and FR-OIF, compared with CF, OIF and the ground truth. The evaluation is carried out in presence of simultaneous temporal fault appearing from UWB and LIO during (20–30) s. Except the fused position obtained from FR-OIF all the estimated positions deviated in presence of faulty measurements.

Figure 12

Variation of Semi-logarithmic root mean square error for position along X-Y-Z obtained from different fusion (CF, OIF, FR-OIF).

Figure 13

Semi-logarithmic root mean square error for orientation represented in Euler angles, obtained from different fusion (CF, OIF, FR-OIF).