Position Estimation

At a glance

As the name suggests, positioning is all about precisely determining a position. In some applications, this is no longer enough: the orientation of a person or object is also becoming important. This is also true in tracking, in which movements over a specific time period are determined. There are different kinds of tracking depending on where the sensor is located: if it is at the edge of the observed area, this is known as outside-in tracking; if it is on the object or person, the term is inside-out tracking.

Fraunhofer IIS offers expertise in both fields, using these modules to calculate what is known as a truth value. With the help of movement classification and statistical filters, movements or positions can be estimated and subsequently smoothed. The more data there is, the better the filter can be stabilized and the more accurate the calculated value.

Sensors

The raw data gathered by the sensors must first be filtered before characteristics can be extracted from it. To this end, the data is sent wirelessly to a computer, where it is interpreted using a deep learning algorithm.

The sensors, which are used in robotics and autonomous driving, provide various inertial measurements (INS/IMU). Data from a sensor fusion, including sensors for speed, acceleration and magnetic field, can provide a very accurate value. The calculation also makes use of well-known infrastructure such as WIFI and Bluetooth. Finally, UWB radio is used to collect additional positioning data and transmit the information. The deep learning algorithm developed in-house by Fraunhofer IIS can calculate the precise position and determine which movements occurred.

Methods

Expand all Close all

Machine & Deep learning

Fraunhofer IIS uses machine learning and deep learning for research and development work in a number of its areas of expertise. Conventional machine learning approaches are used when classifying movements or poses: the algorithm can categorize which movements will occur. Depending on the application, it treats objects, people and animals separately. This kind of movement classification helps avoid errors in calculating a truth value because it recognizes when erroneous sensor data suggests a major divergence from the expected result.

The truth value is also used to stabilize the statistical filters (Bayes filters) with a larger pool of values, making the statistical approach more reliable and accurate.

Statistic filters

To achieve greater accuracy when estimating the absolute truth value, statistical filters are used that incorporate the likelihood of the next “step.” These are mostly Bayes filters, but also Kalman filters or particle filters. This is also how the calculated positions are smoothed, or brought into line, since measurements often diverge slightly from the probable direction of movement. The other way around, filters can be stabilized by continuously repeating the evaluation during use, which causes the data pool for the statistically determined value grow further.

Pedestrians dead reckoning

Another way to measure movement is with pedestrian dead reckoning. This iterative method determines the length and direction of each step after it is taken. The length can be determined by way of the distance from the starting point and the direction by way of the angle. In the case of people, for instance, this method is particularly good at determining the stride length and orientation for pedestrians.

Fusion

It often makes sense to combine different positioning solutions. For instance, while radio is less accurate, it can fill gaps when an optical system fails because the camera is masked. And while optical positioning is cheaper, radio can tell the difference between objects and people – in other words, it can recognize them.

Fraunhofer IIS makes use of these kinds of complementary effects when drawing on its expertise in various positioning methods.

Virtual / Augmented/ Mixed Reality

Human-centered sensor fusion perfects the illusion of the virtual world. The virtual image must align precisely with the movement of the head and body. To this end, it is important to detect even the smallest movements and their direction. Other nonlinear systems such as animals can also be described.

Industry 4.0

Position and direction sensing also makes sense in an industrial context, for instance to determine the orientation of forklift trucks so as to optimize their routes. This can increase productivity in logistics.

Sports

Orientation information and tracking also add value in the world of sports. For instance, they can improve training analysis and optimize training. Overlaying the information they generate onto live TV coverage might also lift the entertainment value to a new level.

Fingerprinting Dataset for Positioning

High precision indoor localization is a key enabler for various tasks in health care, industrial production or networking. While classical time-of-arrival (ToA) approaches fail in none-line-of-sight (NLoS) situations, fingerprinting based methods can achieve high localization accuracies under harsh radio propagation conditions.

UWB and 5G Dataset

We offer...

As a partner in research, Fraunhofer IIS offers competence bundling in the field of machine learning and localization. In this way, our engineers and developers adapt the algorithm to your customer requirements, develop it further and thus work out a solution in dialogue with you.