2.1. Dynamic Bayesian Network Representation
In order to derive the optimal Bayesian estimator, we formulate the positioning problem as a dynamic Bayesian network (DBN). We extend the FootSLAM DBN by a feature map based on our MP model. The resulting network is shown in
Figure 1.
The pedestrian is considered as an integral part of the system, who navigates in a map
that embodies moving constraints through walls, doors, obstacles, and characteristic features like elevator and escalator position. The human relies on visual cues
to identify the environment and forms intentions
where to go next. To reduce setup complexity, we employ no means to observe the human perception and intentions, instead we hitchhike on resulting steps
at time
k with odometry measurements
, e.g., obtained by the ZUPT-aided inertial navigation system (INS) described in [
6]. In addition, we learn the map of the environment that influences the human intentions and restricts the possible human motion. Since the map of walkable areas is learned during the walk, changes in the map will also be learned after a while. It is assumed that those changes happen only rarely.
In the real world, the odometry measurements are affected by measurement noise and correlated errors . describes the pose, and changes in this position from time step to time k are represented by the pedestrian displacement vector .
Additionally, there are specific features or landmarks, which guide the pedestrian when they intend to perform a certain location-related activity such as changing the floor in a multistory building. We propose to mark such locations through feature identifiers that will be stored in a FM .
For incorporating MPs into this system, we interpret
as superimposed displacements from stepped locomotion
and external locomotion through MPs
. This simplification suggests that the platform adds a motion to the pedestrians without preventing them from stepping within the influence area of the MP. With an IMUs mounted on the foot, we can only reliably measure the continuous movement of the platform during stance phases, but we can detect its occurrence by detecting features, that are elaborated in [
27,
28,
32] and reviewed in
Section 2.3. From these features, we can propagate a belief about the platform and mark the occurrences in the new introduced FM
. Given the platform detection output, we can either use the estimated velocity (see
Section 2.3) if available or sample an assumed velocity vector
that superimposes the MP displacement within one update interval.
The underlying thought is to interpret the MPs as a characteristic of the map. Our justification is its temporal stability—in normal operation mode the elevator or escalator moves with manufacture-dependent (factory-dependent) constant (absolute) velocity—and the moving constraints enforced by the car size or width between handrails. In case the platform superimposes a movement on the pedestrian, the external velocity can be perceived by the MPD (
Section 2.3) and an association is made. In case the MP is out-of-order and in rest, no such velocity is detected and no association is triggered.
Our goal is to estimate the hidden states of the MPs at any time instance, which requires to compute the joint posterior
, where
. According to the FastSLAM approach [
10] the joint posterior can be factorized as follows:
which splits the estimation problem into a mapping and a localization problem. In this equation we exploited the fact that given the history of poses
, the transition map
as well as the FM
become conditionally independent from the history of step vectors
, measurements
and
, and errors
.
We can express the right hand side, i.e., the localization problem in Equation (
1), recursively. In accordance with the DBN structure in
Figure 1 and similar to the derivation in [
17], given the relationship
and assuming that human motion is independent from using an MP the recursion formula becomes
In our representation, we assume to be independent from . Since the velocity is reset within each stance phase in the lower UKF, we can assume that this assumption holds. In addition, we assumed that the error is independent from . Due to the fact that there are different error sources the errors are independent from each other. On the one hand, the step length and heading error resulting from the UKF represent the error sources for . On the other hand, either integration of the acceleration sensor biases for calculating the elevator velocity during stance phases (or assuming a predefined velocity in escalators) and additionally heading errors of the FootSLAM algorithm (escalators) are the error sources for .
In the original FootSLAM approach, the map is divided into a regular grid of adjacent uniform hexagonal prisms and the transitions stored are the transitions over hexagonal prism edges. In the proposed extension to FootSLAM, we additionally store the feature properties as additional parameters to the hexagonal prisms. In our implementation, a new feature vector is added that only needs to be stored once per hexagonal prism of the original transition map. The complexity reduced FootSLAM approach presented in [
40] remains the same except that we additionally store the feature vector in each hexagonal prism.
In the following, we look closer at the first term of Equation (
2), which is the pose transition probability and where the transition map as well as the FM play a role (see DBN,
Figure 1). Assuming that the transition map
, as well as the FM
, are statistically independent for simplicity and that the pose and the true step
are only dependent on the previous pose
and on the respective map, we marginalize this term over the transition map
and the FM
and obtain:
where we defined
and
to be the integrals marginalizing over
and
, respectively. Note that both maps are not fully independent, since an MP constraints the possible transitions in the map. On the other hand, the pedestrian may also walk on the platform, so that other directions than the platform direction (e.g., up and down in an elevator) are still possible. The feature vector is restricted to contain the position, velocity and direction of the platform and not the step displacement and heading of the pedestrian. With the assumption of independent maps we can handle both maps independently and are able to calculate the weights separately.
Since the transition map is a set of hexagonal prisms
with
hexagonal prisms, the conditioned probability of the transition map can be expressed as:
In the same way, the FM is a set of hexagonal prisms
with
hexagonal prisms. Note that the FM consists of the same hexagonal prisms as the transition map due to the fact that we rely on the same history of poses. Therefore, the number of hexagonal prisms in the FM is equal to the number of hexagonal prisms in the transition map. The conditioned probability of the FM can be expressed as:
With this assumption, the integral
becomes:
In [
16] it has been shown that the weight update for each particle
m in the original FootSLAM approach is equal to the last particle weight times the integral:
. In a similar derivation the same can be proven for the weight update including both maps—the transition map and the FM, therefore we obtain:
Following this equation, we can handle the weight updates depending on the transition map and depending on the FM separately.