A Tripartite Filter Design for Seamless Pedestrian Navigation using Split Clustering and Tukey Update
Pekka Peltola, Jialin Xiao, Chris Hill, Terry Moore, Nottingham Geospatial Institute, Nottingham University, UK; Fernando Seco, Antonio R. Jimenez, CAR-CSIC, Madrid, Spain.
Seamless behaviour of a navigation application on a mobile device is desired. Infrastructure of the existing technologies, like WiFi, and the ease in deployment of the novel technologies, like Bluetooth low energy, define the measure for the level of achievable seamlessness. Outdoor navigation solutions are already used in everyday mobile navigation. Indoor navigation, on the other hand, is still a challenge. The next step for the applications is to possess the ability to deliver users both of the solutions - in a seamless fashion. In practice, for example, the device can guide the person from outdoor environment into a shop to buy the desired product without any user intervention.
This paper describes a novel tripartite navigation filter design. The filter consists of three blocks and enables plug-and-play style approach on each block of the navigation system design. The sensor subsystem block reformats the raw data into suitable information for the two other blocks. The context engine block helps in defining the best possible navigation filter settings by adjusting the system parameters accordingly. The navigation filter block uses the best navigation method according to the current context and sensor subsystem availability.
The design was used in the creation of a navigation filter implementation which was tested both indoors and outdoors. The walk tests were made in the Jubilee Campus of the University of Nottingham. The campus represents a challenging urban environment with areas susceptible to multipath. Moreover, continuous and adequate positioning solution is maintained when entering a building and when leaving a building.
The filter trials include four preselected technologies and is targeted for pedestrian applications at first. The design is modular and easily extendable to vehicle based navigation. Inertial dead-reckoning tracks the foot and offers a short term solution for both indoor and outdoor environments.
Inertial Measurement Unit (IMU) that was used in the tests was the Microstrain 3DM-GX4-45, which also contains a U-Blox GNSS receiver. The GNSS receiver provides position information outdoors which is fused with the IMU data. Indoors, the Bluetooth low energy and Ultra-wideband sensor subsystem position estimates replace the GNSS in the fusion process.
The fusion is based on availability of the measurements from the sensors. The design takes into account the gaps in the reception and continues running using the available information. The novel filter design processes and fuses the measurement information instantly. C program for live recording of IMU, GNSS, BLE and UWB produces the input files. The filter is implemented in MATLAB. The design is made for real-time applications. Thus implementing the filter on a mobile device is straightforward.
The filter design adapts to the current situation depending on the behavioural movement features, location, the building map and infrastructure information. A context inference engine is used for setting up the navigation filter settings. The navigation filter has 2 states. Depending on the context, the computational power that is used in the solution derivation is automatically adjusted to the required level. The transitions between the states happen, when trigger conditions are in effect.
The filter starts in the particle filter mode. This is computationally heavy but also offers the most robust navigation solution. Depending on the available stand-alone positioning systems' measurements the particles are initialized around the start point with corresponding position distribution. Particles interact with the map walls.
A split clustering enables the particle cloud to be divided into multiple navigation solutions. Split clustering is used with a sparseness threshold that defines the splitting of the particle cluster. Multiple hypotheses are followed.
If the particle cluster characteristics allow, a switch to an error state Kalman filter representation is performed. This trigger is a threshold value of the particles’ positions standard deviation from the particle cluster position solution. The particle cluster is deleted freeing computational power. The Kalman solution is surrounded with a pentagon buffer that consists of five positions around the individual main solution. These five points represent the 3-sigma points of the main solution. In this state, only the pentagon buffer interacts with the map walls releasing computational power, when it is no longer necessary to map match each particle separately.
The individual navigation solution is now represented by a Kalman filter and a pentagon buffer. Interacting with the map the pentagon polygon successfully provides correcting information when the main solution bumps into walls using a collision logic.
The Kalman filter with a pentagon shaped buffer is computationally much lighter than the particle filter and is preferred. In extreme collisions, particle filter needs to be reinitialized. The tripartite filter fails rarely. The system reaches meter-level accuracy in the outdoor trials and sub meter-level accuracy indoors depending on the measurement availability and quality.