13 September 2021 in IEEE Transactions on Instrumentation and Measurement
This article presents a loosely coupled visual/inertial hybrid system for relative attitude determination on a moving platform. The inertial measurements are accomplished by two gyroscopes, of which the slave gyroscope removes the redundant motion of the platform from the total motion sensed by the master gyroscope. To correct integration drift error and compensate for the gyroscope biases, we fuse the inertial and visual measurements fused together through the error-state Kalman filter (ESKF). Furthermore, an improved ESKF is proposed by utilizing the variational Bayesian (VB) inference to adaptively estimate the process and measurement noise covariance. To realize direct estimation of the process noise covariance, a latent variable is introduced and inferred together with the noise covariances and error-state. The practical experimental setup is built up and experiments of two different motion trajectories are conducted. The ESKF, the cubature Kalman filter (CKF), the extended Kalman filter (EKF), the proposed filter, and the other two adaptive filters are implemented for comparison to verify the effectiveness of the proposed filter. The results demonstrate the proposed filter show superiority in reducing both the root mean square error (RMSE) and maximum error (ME) and improving the filter consistency.
17 April 2023 in IEEE Transactions on Instrumentation and Measurement
In stereo phase-shifting profilometry (PSP), the performance of stereo matching algorithm directly determines the result of final reconstruction. Deep learning has been introduced in correspondence retrieval owing to its superiority in solving inverse problems. However, traditional training-based approach requires huge and representative datasets, making it more restrictive for real applications. Inspired by the excellent capability of generalization of Deep Image Prior in untrained image restoration, our work tries to interpret the principle of correspondence retrieval in an image restoration manner, which enables the utilization of Deep Image Prior -based training-free deep network. The inner relevance and invariance both pixel-wise and patch-wise across stereo image pairs are exploited with Deep Image Prior instead of classified prior learned from external dataset, which means that our framework relies merely on an untrained convolutional neural network to accurately retrieve the pixel correspondences without any assumption of system geometry while ensuring the consistency of the reconstruction. It is experimentally demonstrated that the proposed method faithfully realizes correct and dense matching for diverse scenes including both single and multiple objects, competitive or even better performance is achieved against prevailing correspondence retrieval methods.
19 January 2024 in IEEE Sensors Journal
Attitude estimation from unknown corresponding points plays a crucial role in visual tasks involving feature points, but current solutions suffer from slow computation and poor tolerance to noise. In order to address these drawbacks, this paper proposes an IMU-assisted uncertainty-weighted attitude estimation algorithm from non-corresponding points. The algorithm enhances both computation speed and robustness to anisotropic noise. The method begins with an initial pose generator based on pre-integration, optimizing the selection of the current frame’s initial pose using the previous frame’s pose and IMU measurements between them. These strategies enhance calculation speed. Moreover, the feature point extraction method is further optimized, leveraging the covariance matrix to estimate the error associated with the feature points. Consequently, a novel spatial collinearity error function based on SoftAssign is formulated by integrating the Mahalanobis distance. This approach ensures reliable attitude estimation and correspondence determination. Simulation and experimental results demonstrate that the proposed method significantly improves computational speed while achieving superior accuracy and robustness compared to existing methods.
1 February 2024 in Measurement
2 May 2023 in IEEE Access
Image geometric distortion is common in a large field of view optical system. It becomes a technical difficulty for high distortion rate optical systems and high resolution image geometric distortion correction. But, a crucial step in distortion correction is figuring out how to precisely obtain the index coordinates and correction parameters, otherwise known as the lookup table. In this paper, the problems in the traditional lookup table method are analyzed, and an improved lookup table method is proposed. The method is based on the traditional lookup table method with the addition of removing the wrong data of position coordinates in the correction image feature points, calculating the ideal image feature point position coordinates, handling the index coordinate out-of-bounds, and adjusting the output order of index coordinates and correction parameters. The experimental results show that the improved design method is feasible, and improves the accuracy of the correction image, compared with the traditional lookup table, the root mean square error is reduced by 93%, and the average accuracy of the fitted position over the full field of view is reduced by 51.9%, and the maximum deviation is only 0.2917%. For the correction of optical display devices with a wide field of vision, it offers good correction effects.