1. Introduction
With the popularity of intelligent vehicles, positioning and navigation technology is particularly important, because the foremost requirement for all of these safety and assistance systems is an accurate knowledge of the vehicular states (including vehicle position, vehicle velocity, vehicle acceleration, vehicle attitude, etc.) at all times [
1]. High-precision sensors can certainly obtain very accurate vehicular states, but their high prices prevent them from being used in general-priced vehicles. Developing affordable sensors is the main research trend for driving applications [
2]. On the other hand, the sensor should offer continuous and higher-update-rate observations under all circumstances for driving applications. Thus, considering cost-efficiency and high sampling frequency, a low-cost inertial measurement unit (IMU) is a good option and could autonomously provide navigation information. Furthermore, IMUs mostly contain a gyroscope and accelerometer for all three axes which can offer sufficient navigation information about a sudden acceleration or angular and heading change [
3]. However, the accuracy of sensors decreases with the price reduction, and the performance is seriously affected by accumulated bias, drift, and sensor noises [
4]. While an automotive-grade gyroscope typically gives drift performance of 1°/h, a microelectromechanical system (MEMS) gyroscope has a typical performance of 70°/h [
1]. Plainly, low precision is one of the most critical obstacles in the development of a low-cost IMU, which limits its applications such as navigation and guidance [
5]. An approach to design sensors for systems requiring better performance than the one with a single low-accuracy IMU may offer is to fuse the measurements of multiple low-accuracy IMUs to achieve the advantage of complementarity [
6]. This “wisdom of the crowd” design also has, in addition to the improved measurement performance, the benefit of making it possible for sensor fault detection and diagnosis, thereby increasing the integrity and reliability of the system [
6]. Many related studies confirmed the feasibility of this design, including Leszczynski (2014), Martin (2013), Skog (2014), Wahlström (2018), etc. [
6,
7,
8,
9]. However, these studies for improving the accuracy of MEMS IMUs only focused on the device itself.
The global positioning system (GPS) is by far the most widely used high-precision localization method. However, it is further constrained by the non-availability of location during signal outage in dense urban areas [
10,
11]. Additionally, for determining sudden acceleration and braking, the GPS shows very limited success due to the insufficient update rate. Although both the low-cost IMU and GPS have their merits and demerits, they are complementary [
4]. All things considered, one preferable approach is to integrate multiple low-cost inertial measurement units (IMUs) and one GPS receiver for kinematic positioning and navigation in mass-market automotive applications.
The optimal filter design is the key to combining multiple sensors for accuracy improvement, and the adoption of the Kalman filter (KF) is widespread. Conventionally, the Kalman filter based on integration mechanization usually adopts the inertial navigation mechanization to estimate the error states and sensor systematic errors through error measurements on the basis of the aiding sensors [
12,
13,
14]. Its strong dependence on the a priori inertial error model may limit the use of a low-cost IMU, because the time-variant noise model could be highly sensitive to the dynamic excitations and the temperature [
15]. At the same time, the direct Kalman filter, which can accurately reflect the evolution of the real state, was also occasionally discussed and applied [
16,
17,
18]. However, these system models used in Kalman filters are all based on the a priori inertial error model. In other words, in all cases, there is no essential difference, as the IMU measurements are only applied in free inertial navigation calculation and, thus, no measurement updates are performed in Kalman filtering between adjacent measurement epochs [
19]. The severe drift of low-cost IMU systematic errors occurring during GPS outages can easily result in an intolerable free inertial navigation solution between two aiding measurement updates [
20,
21]. To address the realization of a better utilization of measurements from IMUs in Kalman filtering, especially with low-cost IMUs, several studies were carried out. An unconventional Kalman filter was already developed by Wang and Qian [
19], specifically (1) utilizing a three-dimensional (3D) kinematic trajectory model as the central system model, so that the influences of the time-variant errors of low-cost IMUs on the inertial navigation solution are essentially alleviated; (2) allowing the measurements of all sensors, IMU included, directly and independently participating in measurement updates of Kalman filtering, so that the heavy dependence of the inertial navigation mechanization on the IMU measurements in the conventional integration strategy is released [
15,
19]. The unconventionality of this integration strategy mainly embodies in (1) implementation of the direct estimation of the whole-value states (navigational parameters) via sensor measurements instead of the error states via error measurements; (2) the lack of the need to make a distinction between the aiding sensors and the core sensors, since the raw observables (specific forces and angular rates) of IMUs directly participate in measurement updates as the raw outputs from sensors such as a GPS does, thereby implementing true multi-sensor integration.
This research applies the unconventional integration strategy for multiple low-cost IMUs and a GPS-integrated system. There are some characteristics and some corresponding defects from previous 3D kinematic trajectory models. Inevitably, a road vehicle is certain to experience severe or light irregular changes of velocity from time to time, and the changing velocity indicates the presence of acceleration. However, the jerk vector in the kinematic trajectory model is usually treated as process noise. In most cases, the a priori knowledge of vehicle maneuvers is very little. Because the maneuvering process is controlled by human forces, it is difficult to be accurately described with mathematical formulas and can only be artificially approximated under various assumptions. Therefore, a less complex kinematic model cannot realize smooth transitions between different motion statuses for the road vehicle, which leads to an intolerable positioning and navigation solution. One should avoid employing the most complex kinematic model as the core system model throughout [
18]. An appropriate model will not only be able to properly model the vehicle maneuvers, but also exhibit mathematical tractability.
We found that the key problem is how to model the unknown target accelerations. Some mature maneuvering target models were proposed on this issue. So far, the most influential maneuvering target model is the Singer model proposed by Singer in 1970, which attempts to model large-scale maneuvers by assuming a time-correlated input process and incorporating the statistics into the subsequent filter design [
22]. As a global statistical model, the Singer model considers all maneuvering possibilities of targets and can be applied to many types of maneuvering. However, the Singer model is essentially an a priori model, which has some limitations in effectively describing random maneuvers of the target. Based on the Singer model, a series of modifications were developed, including the “current” statistical (CS) model, adaptive current input statistical model (ACISM), jerk model, coordinate turn model, etc. [
23,
24]. The common feature of these models is that the maneuver of the target is regarded as the result of a time-correlated colored noise sequence, rather than a statistically independent white noise sequence. Among them, the ACISM considers the disturbance of maneuvering acceleration caused by environmental noises and other factors, the jerk model is more suitable for describing jerky maneuvering behaviors of highly maneuvering targets (such as aircraft), and the coordinate turn model is mainly applied to turning motions. There is no comparatively great environmental noise (such as atmospheric turbulence) or strong maneuvers in the application of this research. Therefore, the “current” statistical model, which can fulfil the needs of this study, is used to model the system acceleration. As one enhanced part of this manuscript, a more practical 3D kinematic trajectory model based on the “current” statistical Singer acceleration model (CSSAM) is established as the main part of the system equation so that the smooth transitions for the maneuvering vehicle can be realized.
In addition, the significant approach to how the low-cost IMU arrays are fused in this research involves enabling the direct use of measurements for individual IMUs and separately modeling their systematic errors in Kalman filtering, in contrast to most existing approaches working under the supposition of “the common-mode errors of different sensors of the same design” [
7], which means modeling their systematic errors as a group of the shared states in KF from the view of algorithm design. Usually, one created a high-performance artificial or virtual IMU to be equivalent to this IMU array [
8,
25]. In general, those IMU systematic errors are different from run to run and in-run even with the same IMU sensor, which was recognized and largely admitted by the community of IMU manufacturers and users. The centralized filter fusion in Bancroft and Lachapelle (2011) could process the relative position, velocity, and attitude between the IMUs in order not to repeatedly use the GPS measurements once a GPS aiding epoch became available [
25]. However, none of the aforementioned outcomes allowed applying the measurements from each IMU and GPS receiver independently in KF measurement updates. As another enhanced part of this manuscript, measurements and systematic errors of these IMU arrays are individually modeled in Kalman filtering, instead of adopting a set of common shared states for all IMUs. Therefore, the effect of the noises of low-cost IMU raw outputs could be markedly reduced because both the raw data and systematic errors participate in Kalman filtering updates. Moreover, we have the choice of multiple low-cost IMUs and need not stick to the ones of the same design or from the same vendor.
As a matter of fact, given the integration strategy and the system model, various nonlinear filtering candidates (e.g., extended Kalman filter (EKF), unscented Kalman filter (UKF), particle filter (PF), etc.) are available. Nevertheless, this manuscript uses the EKF instead of other nonlinear filtering algorithms because (1) the research does not focus on the filtering techniques, (2) the system model does not present strong nonlinearity, and (3) the EKF algorithm is simple and easy to use in practical engineering.
A more practical system motion model is described in
Section 2. The formulation of the unconventional multi-sensor Kalman filter is presented in
Section 3. The implementation of the Kalman filter is discussed in
Section 4. The data processing results from road tests demonstrate the feasibility by adopting the proposed algorithm model under the unconventional integration strategy. The corresponding numerical results along with analysis are provided in
Section 5. Conclusions and remarks are given in
Section 6.
4. The Implementation of the Kalman Filter
On the basis of the system model proposed above, the KF is structured straightforwardly. The state Equations (22)–(32) and measurement Equations (33)–(39) are separately generalized by the discrete nonlinear system and measurement models as shown below.
where
is the state vector as determined in
Section 3.1,
is the measurement vector as described in
Section 3.3 and
Section 3.4,
and
are nonlinear mathematical functions, and they are constructed from Equations (22)–(32) and Equations (33)–(39), respectively,
and
are coefficient matrices,
is the system input,
,
is the measurement noise vector, and
is the process noise vector including the jerk vector, the derivative of the angular rate vector, and so forth. Specifically,
,
, and
are given as
As is well known, the Kalman filter performs the prediction of the state vector through two information update processes: time update and measurement update. Specifically, the one-step prediction and the variance propagation of the state vector proceed from epoch
to
during the time update.
where
is the mean square error matrix of the state vector, and
is the a priori variance–covariance matrix of the process noise vector.
is the Jacobian matrix of the nonlinear system model in Equation (40), with
(
represent the serial numbers;
represents the
i-th system equation;
represents the
j-th component of state vector
), as follows:
with
Subsequently, the measurement update is performed as follows when the measurements at epoch
are available, thus accomplishing the state estimation:
where
is the Kalman filter gain matrix, And
is the a priori variance–covariance matrix of the measurement noise vector.
is the Jacobian matrix of the nonlinear measurement model in Equation (41), with
(
represent the serial numbers;
represents the
i-th measurement equation;
represents the
j-th component of state vector
), as follows:
with
, wherein,
,
5. Road Test and Results
The proposed algorithm model under the unconventional integration strategy was adopted to process the navigation data from multiple low-cost IMUs and a GPS-integrated system on a land vehicle. Several road tests were performed by our ground-based vehicle navigation system with a Harxon Mini Survey Antenna GPS500 and three IMUs (FOS/05M, ADIS16405BMLZ, and Crossbow IMU440CA), as shown in
Figure 4.
Table 1 shows the performance parameters of the three IMUs. The experimental results are given in this section from one of our road tests. The used dataset was collected in Harbin, China, from which an 8-min data fragment of the data source was selected for demonstration purposes.
Figure 5a reveals the environment where the measurements were made. As can be seen from
Figure 5a, the chosen test environment was an urban highway, with the Songhua River and buildings nearby.
Figure 5b,c depict the trajectory and velocity profile of the test vehicle.
In order to obtain the ground truth as the reference, we equipped a high-grade Fiber-Optic Gyroscope (FOG) inertial navigation system (INS) (gyroscope: constant drift less than 0.01°/h, random noise less than 0.001°/h; accelerometer: constant bias less than 100
, random noise less than 10
) developed by our research group on the test vehicle, as shown in
Figure 4a. Hence, the benchmarks for attitude, velocity, and position could be provided by fusing the measurements of the high-grade INS and the GPS.
5.1. Further Insights into the Unconventional Integration Mechanism
The distinctions between the conventional integration mechanism (left side) and the unconventional counterpart (right side) are shown in
Figure 6: (1) the embedment of the prediction of the kinematic states, (2) the introduction of IMU measurements, and (3) the navigation parameters directly used in the state vector.
The acceleration vector and angular rate vector are only regarded as inputs of the mechanization in the conventional error state-based Kalman filter, while the realistic measurements of the IMU directly participate in the measurement update process in the unconventional Kalman filter. In principle, the Kalman filter is equivalent to a sequential least square with a time-variant state vector and the process noises [
30]. That is to say, the system model is composed of a group of virtual measurements for the state vector and reflects the connections between the state vectors from epoch to epoch. These extra virtual measurements mean that the measurement redundancies in the unconventional Kalman filter for the body acceleration vector
and the body angular rate vector
are evidently better than those in the conventional Kalman filter. Therefore, the accuracy of
and
in the unconventional KF for multiple low-cost IMUs and a GPS-integrated system will be undoubtedly improved because they are not only predicted but also measured. The prediction for
and
in the system model can be used as a rigorous reference to check on the performance of the IMU without increasing the complexity of the filtering structure. Furthermore, the system equations of the unconventional KF can act as the dynamic constraints for the navigation parameters, for example, assuming that
and/or
.
Generally speaking, the most significant feature of the unconventional integration mechanism lies in the improvement of the overall measurement redundancy of the system through the system model based on the 3D kinematic trajectory model.
As discussed above, it is expected that the accuracy of the navigation solutions shall be improved by using the rigorous trajectory model as the system model in the unconventional KF. As the components of the state vector, the acceleration and angular rate vectors in the body frame (
) will also profit from the unconventional KF time updates for the same reason. For better visual effects,
Figure 7 and
Figure 8 show the comparisons between the raw IMU outputs and filtered signals.
As can be seen from
Figure 7, the overall variation of the IMU raw angular rates is slightly larger than that of the filtered angular rates. However, it is apparent from
Figure 8 that the overall fluctuation of the filtered accelerations is significantly lower than that of the IMU raw specific forces. The improvement is due to the introduction of the novel system model in the unconventional Kalman filter. It is worth mentioning that the raw output of the accelerometer in the vertical direction includes acceleration due to gravity, while the corresponding filtered acceleration does not. That is why there is a large difference (about 10
) between the raw and filtered acceleration values along the vertical axis in
Figure 8; here, we only compare the fluctuating ranges of the two curves.
The comparison results from
Figure 7 and
Figure 8 confirm the fact that the influences of the IMU measurement noises on the final navigation solutions are effectively mitigated due to the participation of the IMU outputs in the KF measurement updates.
5.2. Verification of the Proposed Algorithm Model under the Unconventional Integration Strategy
Figure 9,
Figure 10,
Figure 11 and
Figure 12 exhibit the solution accuracies for the kinematic trajectory parameters and attitude using the proposed algorithm model under the unconventional integration strategy.
Figure 9 shows the solution accuracy comparisons for 3D position using the basic 3D kinematic trajectory model or the practical model based on the “current” statistical Singer acceleration model (CSSAM). The overall 3D position accuracy using the proposed algorithm model is under 5
, which is a great improvement compared with the accuracy (10
) using the basic one.
Figure 10 shows that the estimation errors for the velocity state vector in the three axial directions of the body frame using the proposed algorithm model are within
,
, and
, respectively. The velocity error in the direction of travel is obviously reduced compared with that using the basic model.
Figure 11 shows that the estimation errors for acceleration state vector in the three axial directions of the body frame are within
,
, and
, respectively. As can be seen, the accuracy of the acceleration estimation is apparently improved by using the proposed algorithm model. As shown in
Figure 12, the accuracies for attitude (pitch, roll, and heading) are within
,
, and
, respectively. It is to be observed that the pitch angle has an apparent fluctuation from 3–4.5 min. The trajectory curve in
Figure 5b illustrates that the vehicle experienced ups and downs in the test duration, which could affect the estimation of attitude angles (especially for pitch angle). It also indicates that the attitude estimation model still needs improvement so that it can be applied to more complex motion forms.
Apparently, the navigation parameters during the 8-min vehicle motion were estimated within acceptable ranges by adopting the proposed algorithm model under the unconventional integration strategy. Furthermore, the reason why the 3D position error using the basic trajectory model did not diverge is that the output of the central accelerometer with real-time calibration was utilized to regulate the acceleration estimation error when significant acceleration maneuvers occurred.
5.3. IMU Systematic Error Estimation
Since the systematic errors of these IMU arrays, i.e., biases and scale factor errors of gyroscopes and accelerometers, are individually modeled in Kalman filtering, the systematic error estimations of each IMU can be obtained separately, as illustrated in
Figure 13,
Figure 14,
Figure 15 and
Figure 16. As an example, this manuscript only shows the systematic errors of the central IMU due to space limitations. The common existing approach under the conventional inertial navigation mechanization adopts a set of common shared errors for all IMUs and, sometimes, these error parameters are from initial calibration results or technical specifications [
7,
28]. However, in fact, the a priori error model defined for a static low-cost MEMS inertial sensor needs to be checked and compensated for in the dynamic working environment as the vibration on a low-cost IMU might cause significant changes in its scale factors and noise level compared to those in the static case [
31].
As can be seen from
Figure 13,
Figure 14,
Figure 15 and
Figure 16, in contrast to the a priori constant systematic error with the common existing approach, the systematic error estimations with the individual modeling method vary with time, which conforms better with the real situation. In other words, the individual modeling method under the unconventional integration strategy can adjust the systematic error estimations of each IMU according to the real-time measurements from each IMU. This technique can firstly verify if those IMUs really share the same systematic errors quantitatively, even if they are the common errors physically; then, they can be modeled either separately or combined, laying the groundwork for future research such as auto calibration and fault detection. While applying the individual modeling method for the IMU array, in spite of its characteristic properties, one must confront the following problem: high computation load caused by high-rate measurement updates in the Kalman filter, e.g., with IMUs whose measurement rate may be at 100 Hz. Although the modern computation capability is considerably improved, one cannot stand such a high measurement update rate, especially one so unreasonable, which renders it useless in real time. How to appropriately reduce the high-rate KF measurement updates without compromising the valuable information embedded in the high-rate measurements will become a topic for further study.