PLV-IEKF: Consistent Visual-Inertial Odometry using Points, Lines, and Vanishing Points

IEKF [19] is proposed to address the filter inconsistency problem, which has been widely researched in a range of literature [18, 20, 21, 22, 23, 24, 25]. The IEKF model has been further applied to MSCKF [12, 14, 26] and UKF [15]. Previous works [14], [27] have coupled the landmark uncertainty with the camera pose, while other works adopt the simple addictive error [26]. Though both strategies can maintain the consistency, the relationship between two kinds of error definition has not been investigated.

A couple of works have combined point and line features. Among optimization-based methods, [9] jointly minimizes the IMU pre-integration constraints together with the point and line re-projection errors. [28] further modifies the LSD [29] algorithm to have a better real-time performance based on [9]. By utilizing the structural regularity, Lee et al. introduce parallel lines’ constraints into point-line-based VIO and present a novel structural line triangulation method making full use of the prior information [13]. Among filter-based methods, the issue of point-line VIO inconsistency is addressed and improved by observability-constrained techniques [30] and the IEKF framework [12]. [31] extends a new line parameterization for processing the line observations in a rolling shutter camera setting. Zheng et al. present Trifo-VIO which incorporates point and line feature measurements and formulate loop closure as EKF updates [32]. [33] investigates two line triangulation methods and reveals three degenerate motions which cause triangulation failures.

In man-made environments, geometric information like vanishing points can also boost the robustness of VIO and Simultaneous Localization and Mapping (SLAM) [34, 35, 36, 37]. Zhang et al. provide a solution to the loop closure problem leveraging vanishing points in line-based SLAM [38]. [39] proposes an efficient line classification method for detecting vanishing points and keeps the vanishing points in the VIO state for a long tracking. Xu et al. design a point-line-based VIO system where vanishing points are used to recognize and classify structural lines [40]. In [10], the vanishing point cost function is added into the optimization-based VIO without the constraints like the Manhattan world assumption [41]. However, these works have not focused on consistency about vanishing point measurements.

In the visual-inertial navigation system, the defined state is given by:

where $\mathbf{p}_{f}$ is the landmark. Unlike the conventional error in standard EKF, the right invariant filter employs a new uncertainty representation:

where $\mathbf{J}_{l}$ is the left Jacobian for $SO3$ Lie group, and $\xi=\begin{bmatrix}\xi_{\theta}^{T}&\xi_{v}^{T}&\xi_{p}^{T}&\xi_{b_{g}}^{T}&\xi_{b_{a}}^{T}&\xi_{p_{f}}^{T}\end{bmatrix}^{T}$ is the state error vector. Note that apart from robot velocity and position, the landmark’s uncertainty is also coupled with the rotation state. This nonlinear error allows the system to obtain a better performance on consistency.

There exist many representation models for a 3D line, and two representations are focused on in our paper. One is the conventional Plucker coordinate $\mathcal{L}=\begin{bmatrix}\mathbf{n}^{T}&\mathbf{d}^{T}\end{bmatrix}^{T}\in\mathbb{R}^{6}$, where $\mathbf{n}\in\mathbb{R}^{3}$ is the normal vector of the plane determined by the line and the coordinate origin, and $\mathbf{d}\in\mathbb{R}^{3}$ is the line direction vector [9]. However, this representation has redundant parameters since the degrees of freedom for a line should be four, so the minimal representation $\mathbf{o}=\begin{bmatrix}\mathbf{\psi}^{T}&\phi\end{bmatrix}^{T}$ is more suitable in line optimization:

where $\lfloor\cdot\rfloor$ denotes the skew symmetric matrix.

The IMU continuous kinematic model is given by:

where ${}^{I}\boldsymbol{\omega}$ and ${}^{I}\mathbf{a}$ are the angular velocity and linear acceleration velocity, respectively. $\mathbf{n}_{wg}$ and $\mathbf{n}_{wa}$ are Gaussian noises. After linearization, the following propagation equation is obtained:

where $\mathbf{n}_{I}=\begin{bmatrix}\mathbf{n}_{g}^{T}&\mathbf{n}_{wg}^{T}&\mathbf{n}_{a}^{T}&\mathbf{n}_{wa}^{T}\end{bmatrix}^{T}$ is the process noise. $\mathbf{F}$ is the continuous state transition matrix and $\mathbf{G}$ is the input noise Jacobian.

When the position of the point feature is initialized, the normalized coordinate $\mathbf{z}$ of the feature point ${}^{G}\mathbf{p}_{f}$ is obtained:

where $\pi$ is the projection function. The corresponding error vector of $\mathbf{X}$ is denoted as $\mathbf{\xi}=\begin{bmatrix}\mathbf{\xi}_{I}^{T}&\mathbf{\xi}_{C1}^{T}&...&\mathbf{\xi}_{Cn}^{T}\end{bmatrix}^{T}\in\mathbb{R}^{15+6n}$. Therefore, the linearized measurement model can be obtained as below:

where Jacobians $\mathbf{H}_{x}$ and $\mathbf{H}_{f}$ are given by (25) in Appendix. Genrally, ${}^{G}\tilde{\mathbf{p}}_{f}$ follows an invariant error definition like (2):

where $\xi_{\theta}^{i}$ is the orientation error for $\mathbf{X}_{C_{i}}$, and $C_{i}$ is the camera pose that captures the landmark first. However, a proposition is given to reveal the equivalent relationship between the right invariant form and the conventional error form, i.e. ${}^{G}\tilde{\mathbf{p}}_{f}={}^{G}\mathbf{p}_{f}-{}^{G}\hat{\mathbf{p}}_{f}$.

$\mathbf{Proposition}$ $\mathbf{1.}$ Given the same estimated landmark, the measurement model for points is equivalent for the invariant error and additive error.

Proof: See Appendix.

Remark: The equivalent measurement model indicates the same consistency property. However, the additive error definition is more convenient to implement.

As shown in Fig. 2, $\mathbf{I}_{L}$ represents the projection of a straight line $\mathcal{L}$ on the normalized image plane, so the error term is defined as the distance between the endpoints on the line in an image and the estimated 2D line:

where $\mathbf{l}=\begin{bmatrix}l_{1}&l_{2}&l_{3}\end{bmatrix}^{T}$ is the re-projected line on the normalized plane, $\mathbf{p}_{s}=[u_{s}\quad v_{s}\quad 1]^{T}$ and $\mathbf{p}_{e}=[u_{e}\quad v_{e}\quad 1]^{T}$ are the endpoint measurements of the line on the normalized plane.

The line error for the orthonormal representation is classified into globally-defined error and locally-defined error as below:

Thus, a conclusion on these two error forms can be given:

$\mathbf{Proposition}$ $\mathbf{2.}$ Given the same estimated line, the measurement models for lines are equivalent for the globally-defined error and locally-defined error.

Proof: The proof is similar to that of $\mathbf{Proposition}$ 1.

Remark: Conventional line-based VIO usually chooses the locally-defined error [9] while IEKF-based VIO [12] may choose the global one. Since the line feature is not incorporated in the state variables, both errors are acceptable in the algorithm implementation.

Therefore, the linearized measurement model following a globally-defined error is given by:

where $\mathbf{J}_{L}$ is the Jacobian of line projection, $l_{1,2}=\sqrt{l_{1}^{2}+l_{2}^{2}}$.

In the man-made environment, lines can be classified into structural lines with VP measurements and non-structural lines [10]. For a structural line, when its corresponding vanishing point is calculated, a VP measurement residual is obtained:

where $\mathbf{r}_{v}$ and $\mathbf{p}_{v}$ represent the vanishing point residual and the vanishing point observation, respectively. ${}^{C}\mathbf{d}=\begin{bmatrix}d_{1}&d_{2}&d_{3}\end{bmatrix}^{T}$ is the line direction vector in the camera frame, which is orthogonal to the normal vector ${}^{C}\mathbf{n}$ which is used as the line measurement. In this sense, VP measurement is complementary to line measurement and fully exploits the line geometric information. The Jacobian with respect to the vanishing point is computed as follows:

where the Jacobian items are similar to those in (14).

When a new frame comes, line segments are detected by the LSD detector. For each line, the descriptor LBD is calculated for line tracking. Some outlier rejection strategies are employed such as the limited distance between the endpoints of matched lines and the limited angle difference between the line directions. As for VP detection, we use the efficient and robust two-line based VP detection algorithm [42], and the observed lines with the detected VPs will be considered as the structural lines and used in the aforementioned line measurement model and VP measurement model.

The strategy of line estimation is similar to that of point estimation in MSCKF. When a non-structural line is no longer tracked by the current camera frame, it will be triangulated by the dual Plucker matrix method [9]. Then we optimize it by the following cost function:

where $\mathcal{L}_{n}$ are the sets of non-structural line measurements in the sliding window, $\mathbf{l}_{i}$ is the $i$-th line measurement. After the structural line is triangulated, we add a VP cost item for a better estimation:

where $\mathcal{L}_{s}$ are the sets of structural line measurement, $\mathbf{p}_{v,i}$ is the $i$-th vanishing point measurement.

In the simulation test, we assume that a car equipped with a 100Hz IMU and a 10Hz camera loops ten times along a circle with a radius of 6m. A total of 200 points and 140 lines are scattered around the real trajectory, as shown in Fig. 3(a). The points are generated along the inside cylinder wall with a radius of 5 meters and outside cylinder walls with a radius of 7 meters. And the lines are generated along a square wall with a length of 14 meters. The camera captures landmarks in the field of view of the camera within a range of 20 meters. We generate sensor measurements with noise according to the trajectory and landmarks. The initial noise matrix is set as $diag(0.008^{2}\mathbf{I}_{3},0.0004^{2}\mathbf{I}_{3},0.01^{2}\mathbf{I}_{3},0.003^{2}\mathbf{I}_{3})$, and all the feature measurement noise is set as 1 pixel.

Four algorithms are compared in the simulation: standard point based MSCKF (MSCKF), standard point based IEKF, point-line-VP based MSCKF (PLV-MSCKF), and point-line-VP based IEKF (PLV-IEKF). The sliding window size is set as 20 and we only use the landmarks that are captured more than 5 times by the camera for robust estimation. The pose accuracy indicator Root Mean Squared Error (RMSE) and the consistency indicator Averaged Normalized Estimation Error Squared (ANEES) are reported in Fig. 3(b) and (c). It is evident that NEES of IEKF and PLV-IEKF is closer to the ideal NEES (one) than that of MSCKF family, indicating better consistency. Meanwhile, PLV-IEKF achieves a better pose estimation than IEKF with additional line and VP measurements.

To validate the effect of our proposed method, we conduct a real-world experiment on EuRoC Dataset [43], which exhibits varying levels of motion blur and textureless regions [33]. V2_03_difficult sequence is not chosen due to a number of missing camera frames which affect the feature tracking. We compare our algorithm with other state-of-the-art algorithms such as MSCKF [44], IEKF [27], VINS-Mono [45] and PL-VINS [28]. PLV-MSCKF and point-line based invariant IEKF (PL-IEKF) are also implemented for a detailed comparison. For all of the algorithms, we extract no less than 100 point features and 30 line features for robust estimation and disable loop detection for a fair comparison. The sliding window size for these filter-based VIO is set as 20. An example of tracked features in MH_03_medium is shown in Fig. 4. RMSE results are reported in Table I, and the trajectory sample is shown in Fig. 5. It is evident that in most cases using lines and vanishing points can achieve a better pose accuracy, and PLV-IEKF has a better performance than PLV-MSCKF because of better consistency.

It is also observed in Table I that PLV-IEKF performs better in Machine Hall (MH) sequences than PL-IEKF since there exist more structural features in the environment, where vanishing point measurements can improve the line estimation, even in challenging sequences like MH_04_difficult sequence with illumination changing and textureless scenarios.

Table II shows the backend processing time on MH_01_easy sequence for different algorithms, among which the proposed PLV-IEKF requires slightly more time than other filter-based point-based and point-line-based methods, but much less time than VINS-Mono and PL-VINS. It demonstrates that PL-IEKF can achieve competitive and higher time efficiency compared with optimization-based methods.

To prove Proposition 1, we assume that only one point-type landmark in the current sliding window is observed by $n$ camera poses. When the landmark error is additive, the Jacobian for VIO state is given by:

where $\mathbf{J}_{\pi}{}=\begin{bmatrix}\mathbf{J}_{p1}&\mathbf{J}_{p2}&\cdots&\mathbf{J}_{pn}\end{bmatrix}$, $\mathbf{J}_{pi}$ is the Jacobian $\mathbf{J}_{p}$ for the $i$-th camera pose. On the other hand, the Jacobian regarding an invariant form is calculated as follows:

Note that the Jacobians with respect to the landmark for both forms are the same:

which shares the same left null space matrix $\mathbf{A}$. We calculate the difference between the final Jacobians:

Therefore the final Jacobians are equivalent, so is the final projected measurement noise $\mathbf{n}_{o}=\mathbf{A}^{T}\mathbf{n}$.