Infrastructure independent pedestrian localization using dead reckoning and particle filter

Although there is a ubiquitous technology for outdoor navigation based on Global Navigation Satellite System (GNSS), it's not the case for indoor positioning. Advancements in Microelectromechanical System (MEMS) technology made it possible to utilize sensors built-in smartphones in order to be used for inertial positioning. By the increase in use of smartphones, the use of navigation systems for personal use is dramatically increased as well. While there isn't a standard method accepted by biggest smartphone operating system developers, there is one thing clear that benefiting from Inertial Measurement Unit (IMU) for dead reckoning is crucial as it is for outdoor navigation. Pedestrian indoor localization applications vary greatly and most of them require additional infrastructure to be installed. However, there are certain conditions either prevent an infrastructure to be installed or installed infrastructure to be avoided. Pedestrian dead reckoning is one of the processing applications of inertial signals for indoor personal tracking despite the fact that there are challenges to overcome. While there are number of indoor positioning methods, it's very common to utilize those methods in combinations for improved results. Our work presents an application of device based localization. Particularly, device based localization might refer an offline process but it can either be offline or online. There are quite number of existing technologies that have been used to provide indoor positioning, dead reckoning is a relative positioning technique that basically provides next position of an object based on previous position, and it's used for pedestrian localization in particular. The technique used for positioning people is called as pedestrian dead reckoning and it uses smartphone's inertial measurement sensors for this regard. It consists of three stages: Step detection, which is achieved using the accelerometer sensor, stride length estimation, which is an estimation based on step frequency, and heading estimation, which is obtained by using multiple sensors including gyroscope and magnetometer. Because of its sequential methodology, biggest challenge in dead reckoning is to eliminate the effect of drift error accumulated over time. Finally, the position vector based on previous position vector and new measurements is obtained. Propagating the position from PDR method results a single coordinate per step. It includes drift and other errors causing inaccuracy. The problem of positioning requires a more complex model because the uncertainty it includes turns the output into unreliable. Rather than accepting the deterministic output, it's possible to use mathematical methods to reduce errors, improve accuracy and decrease the level of uncertainty for feasibility even for short running periods of time. Hence, positioning requires estimation of system state, changing over time, using a sequence of noisy measurements. The context of filtering in localization area pertains to sensor fusing, error avoiding and scheme matching. The elimination of inherent uncertainty of the sensor data as well as map matching requirements are the top two aspects of filtering. Bayesian filters are basically used to probabilistically estimate the state of dynamic systems where the dynamic system is the pedestrian and the state is the related position vector. In particular, Particle filter, which is a discrete method where beliefs are represented as weighted samples and it refers to arbitrary probability densities, is used due to its advantages on map matching problems. A probability distribution, which indicates estimated position sequentially with particles moving in regard to the sensor measurements, is obtained by propagation of particles, correcting and eliminating particles and resampling to prevent degeneration at every step. In conclusion, available methods for indoor localization are investigated and the theory behind a system based on pedestrian dead reckoning and particle filter explained in detail along with the utilization information of smartphone sensors. Following the analysis, we implemented a system for pedestrian localization based on dead reckoning and particle filter. According to the evaluation of results the system is found to be enough accurate only for short periods of time and is best to use it in conjunction with an absolute localization method for consumer-grade applications and real-life localization requirements.