Localización de personas mediante sensores inercialesy su fusión con otras tecnologías

Abstract

The popularization of smartphones and the improvements (and cost reduction) of Global Navigation Satellite Systems (GNSS) have promoted the development of many positioning systems. These are able to easily locate a person in a map, but due to the limits of the GNSS technology, in indoor environments, without line of sight to satellites, there is a significant increase in the positioning error. This can be overcome using different techniques to estimate the position of the person according to the environment and the sensors and signals available. Some of the techniques are: inertial navigation and PDR using an IMU, beacon based systems (UWB, ultrasound, WiFi, Bluetooth, etc.), signals of opportunity (TV, mobile networks, light, magnetic fingerprints, etc.), image recognition (geo tags, visual navigation, stereo image analysis, etc.), environmental variables (magnetic field, atmospheric pressure, temperature, etc.) among others. This thesis focus in the positioning problem in indoor environments using inertial sensors and sensor fusion with external measurements to improve the estimation and to limit the drift. The positioning algorithm is divided into three parts, the estimation of the relative position using Pedestrian Dead Reckoning (PDR), the sensor fusion scheme, that allows the use of the information from multiple sensors, and the external measures (like the Received Signal Strength from WiFi access points, ranges to UWB devices, GNSS measurements, etc.) used to limit the drift in the estimation. Each part is studied, proposing improvements to reduce the error level. In this research PDR was implemented based on the measurements of an IMU in the foot of a person, generating an Inertial Navigation System with zero velocity updates during the stance phases of the walk pattern. The algorithm is improved modifying the stance detection to filter the output using the median of a delayed window sample. The new stance detection is tested with real IMU measurements and synthetic signals, showing that the method avoids false detections and improves the initial and final points of the stace phase. The position and orientation of the person is usually estimated using an Extended Kalman filter (EKF), but as a way to improve the propagation and corrections of the non lineal states this thesis proposes the use of an Unscented Kalman filter (UKF). Both filters’ estimations were tested using real and synthetic data, were it is observed that the estimation improves with the UKF, at the cost of a small increase in the computing time. The yaw state in a PDR estimation is usually non observable with only ZUPT measurements, therefore a common solution is the use of magnetic field measurements. In indoor environments the field is affected by the metallic structures of the building, and its direct use introduces errors in the estimation, therefore a measurement of the rotation of the magnetic field according to the gyroscope measurements is proposed. This Magnetic Angular Rate Update (MARU) reduces the rate of growth of the heading error from lineal with time to lineal to the squared root of the time. To reduce the drift of PDR, it was fused with external measurements using two proposed schemes, a constraint filter that limits the distance between two estimations and a two level estimation using a particle filter. The constraint filter modifies the pdf of the estimations to eliminate the probabilities where the estimations of the position of two sensors in the body are too far away from each other. The constraint is implemented as a correction of the mean and covariance of the states and when tested fusing foot-mounted PDR, with the estimation of an IMU in the head with UWB position updates, it limited the drift of PDR and reduced the error level of the IMU-UWB system. The second scheme used was a two level estimation based on a particle filter (high level) that propagated the particles using the information from a PDR estimation (low level) and used the external measurements to update the weights. It was observed that the real error of the estimation grows faster than the particle dispersion, therefore a new state was added to the particles to estimate the bias of the turn rate as a way to model the effect of the bias random walk in the gyroscope. The theoretical and simulated results showed a particle dispersion closer to the real error growth. The use of a particle filter allows the implementation of a wider variety of measurements using only the observation function and the error distribution, therefore several measurements from WiFi, RFID, UWB or ZigBee were fused with the PDR estimation . The positioning systems used presented errors of approximately 5 m (90 % of the time), while PDR had a growing with time/distance error, that after fused was able to provide a 2 m error (90 % of the time). Finally, the map of the building was used to correct the estimation according to the walkable areas of the estructure (map matching). To achieve this, the hypothesis (particles) that crossed walls were eliminated. The algorithm was optimized to use only the walls of the room the particle is in, and it was implemented using Matrix operations to be able to run in real time in MATLAB. Using real signals it was proven that the algorithm was capable of locating himself using only the map of the building, the PDR information (initial position was not provided) and a non symmetrical path. The error level will depend on the map, but for our experiment it was observed to be around 1 m. The map matching algorithm when fused with the use of RF information showed a significantly faster convergence, a reduced error level and requires fewer particles to avoid deprivation (therefore it will be faster).