Estimote Indoor Location Example Essay

1. Introduction

According to a 2012 study commissioned by the Alzheimer’s Society of Canada, 747,000 Canadians have some type of cognitive impairment, including dementia, and this number is expected to double by 2031. People with dementia experience challenges with daily activities (e.g., cooking meals, ironing, taking medication, personal care), such as misplacing materials and failing to complete tasks in the right sequence. Having accurate information about an older adult’s daily activities, and the patterns of these activities, can provide rich information on his/her abilities and capacity for functional independence. Major deviations in daily patterns should likely be considered as indicators of a person’s physical, cognitive and/or mental decline. Having such information could alert caregivers of potentially risky events and the need for additional support.

The advancing wave of Internet-of-Things technologies holds immense promise for enabling such data collection and analysis and for delivering appropriate support. In the project, we have been developing a sensor-based platform for non-intrusively monitoring people at home, analyzing the collected data to extract information about the occupants’ activities, simulating the extracted information in a 3D virtual world, and generating recommendations—for themselves and their caregivers. To meet the non-obtrusiveness requirement of our platform, we have excluded from any image and video capture devices. Of course, for the sake of reconstructing the ground truth via manual annotation, the experiments we carried out also included video cameras. However, in a production-environment deployment, no cameras would be used.

This hardware–software platform has been installed in the simulation space—a dedicated teaching-and-research space in the University of Alberta’s Edmonton Clinic Health Academy (ECHA). The is a fully functional apartment with one bedroom, a bathroom, and an open kitchen-and-living space. Infused into the apartment and its furnishings are sensors that record a variety of environmental variables (i.e., levels of light and sound, temperature, and humidity) as well as the activities of the occupant(s) (i.e., their motion and use of furniture, cabinetry, and appliances). The data acquired by these sensors is transmitted and analyzed into a central cloud-based repository. The has recently been redesigned to include Bluetooth Low-Energy (BLE) beacons attached to different objects in the apartment. The occupants can be provided with a smartphone, running a background service that collects, and transmits to the platform, signal-strength measurements from any nearby BLE beacons. These two types of data sources—sensors and beacons—are used to infer the occupants’ locations at each point in time. The server generates textual reports, spatial visualizations, and 3D virtual-world simulations for the inferred movements and activities of every occupant. In addition, it can potentially generate alerts for special incidents, which can be sent to the occupants’ caregivers, or anyone of their choice.

In our previous work, we investigated the trade-offs between the accuracy of the location-estimation process for one occupant, based on PIR (Pyroelectric or “Passive” Infrared) sensors only vs. the overall cost of the sensor installation [1]. Next, we studied how the use of RFIDs in addition to PIRs could be exploited to enable location recognition for multiple occupants [2]. In this paper, we report the results of our recent study on the relative effectiveness of motion sensors and BLE beacons for accurate location estimation for multiple occupants.

Multi-person location estimation, as a first step to activity recognition, is a challenging problem and has received relatively little attention in the literature. This is partly due to the implicit assumption that if the subjects carry with them active (hence, eponymous) devices, each person can be localized independently, in isolation of the rest; hence, any method for estimating the location of a single individual is assumed to generalize to multiple individuals. However, the situation is drastically different when one (or more) of the subjects do not systematically carry/wear such a device, either because they cannot, they do not want to, or they simply forget to—typical of many care-delivery scenarios. Estimating the location of an individual does not yield the same results when applied to a scenario when the individual is alone vs. when the individual is one among many in the observed space. For example, the radio frequency (RF) propagation environment in a given space varies over time because of the dynamics of human and object placement within that space. In fact, [3] has utilized the impact of humans on the RF environment to estimate the locations of multiple subjects, based on models of how the fingerprints of radio signal strength indicators (RSSIs) change in the presence of people in the space. Nevertheless, this method requires a large set of transmitters and receivers to cover the entire area in each room and the placement of the transmitters/receivers needs to be exact to ensure that they are in line of sight (LoS). We address more of the related work in the next section, noting that our assumptions align closer with those of [4] where individuals carry smartphones, with the notable difference that we allow one of the individuals to not wear or carry any identifying device.

Our own previous work on this problem [2] targeted the development of a method using RFID readers embedded in the environment and wearable passive RFID tags. Such an approach is limited in terms of practicality because RFID readers today—especially if endowed with large antennas to attain reasonable range—are difficult to embed in everyday surroundings without expensive retrofitting of the space (and frequently violating the aesthetics of “home” environments). The passive RFID tags also have to be embedded in everyday items (e.g., clothing), and hence the reliability is generally compromised unless specially treated to cope with washing and other everyday wear-and-tear.

In this paper we focus on (a) the fusion of data collected by PIR motion sensors with data collected from tiny BLE beacons attached with simple adhesive glue on surfaces around the home and collected through an application running on the occupants’ Android smartphones. We then (b) evaluate the effectiveness of our method through an empirical study in the , exploring caregiver scenarios where one individual does not wear a device nor carries a smartphone, while the second (typically the caregiver) carries such a device. In addition to its applicability to realistic care-giving scenarios, the main advantage of the technique described here is that the location of the two individuals can be accurately determined.

The rest of this paper is organized as follows. Section 2 places our work in the context of the most recent related work in the field of localization and activity recognition. Section 3 shows the architecture of our system and our location estimation method. Section 4 outlines our experimental methodology and results. At the end, Section 5 concludes the paper with a summary of our findings.

2. Related Work

Over the past decade, the area of indoor location estimation has resulted in many proposals and research findings, with varying degrees of applicability in real environments. One strategy to solving this problem is based on RSSI fingerprinting. RSSI readings for specific points in the space are collected in a database of readings at known locations; at run time, when a new RSSI reading is received, it is compared to the dataset and the “nearest” point (according to some definition of distance) is selected as the likely location. For example, [5] introduced a fingerprinting system utilizing passive RFID tags and four RFID readers and used a k-nearest-neighbor (kNN) method to select the nearest points to the received signal when localizing. In a building with two bedrooms and with the space logically divided in a grid-like fashion of cells of size  m, their method achieved a reported accuracy as 96% when localizing at the granularity of a grid cell. Typical of fingerprinting methods, it requires prior RSSI data collection, which is a task sensitive to the environment that needs to be repeated should the environment change in ways that impact the radio frequency propagation (e.g., when furniture is added/removed/moved).

Another school of thought pays attention to the kinematics and relates the generation of the location estimates with the direction of movement of the individuals. An example is the pedestrian dead-reckoning (PDR) algorithm [6], where new location estimates are based on previous known locations. For example, in [7], a PDR was proposed with WiFi and iBeacon signals, used to calibrate the drifting of the PDR algorithm by converting their RSS values to meters via a path-loss model. This family of methods requires a fall-back scheme for estimating locations in the absence of previous estimates in two cases: (a) when the initial location needs to be established, and (b) when sufficient error has accumulated, based on the estimates following the kinematics, such that a “re-initialization” of the estimation needs to take place. While we take some measures to consider the kinematic behavior of the individuals, we do not rely on it, as the activities in which an individual is engaged in a small indoor space call for frequent changes of direction and speed, and some tasks are fundamentally unsuitable for dead-reckoning approaches (e.g., broom sweeping). In another PDR approach in [8], the authors used WiFi fingerprints to calibrate the PDR error after time, and they performed their experiments when the subject was walking in a path and used the location estimation approach to track the subject. In [9], an RFID-based indoor location estimation is proposed for the elderly living alone, which uses both RSSI for localizing the subject and fuses it with a PDR using accelerometer data to step and direction detection to increase the accuracy.

We hasten to add that in the IPIN 2016 offline competition [10], the best team introduced a PDR with RSS fingerprinting for the initial position with accuracy of 2 m for a single individual in one of the spaces considered. The other four best teams performed RSSI fingerprinting, PDR, and MAC address filtering and fingerprinting with less accurate results.

In [11], PDR is used with WiFi and iBeacon fingerprinting: the iBeacon is used only where the WiFi signal is not strong enough. Similar to profiling, methods that use path-loss models rely on a model-configuration step, specific to the building and the environment where they are deployed, and changes to the environment require a re-computation of the path-loss model parameters to preserve accuracy.

The approach we discuss in this paper considers only RSSI values higher (stronger) than dBm. The choice for this threshold comes from our previous work [12], and reflects situations where the subject is very close (approximately within one meter) of the Estimote beacons we use. In this fashion, the RSSI values only matter when they are strong enough to act as a proximity sensors rather than as a model for distance calculation.

A self-calibrated system is proposed in [13], where the sensors (smartphones in this case) communicate with each other to determine first their locations relative to each other, and subsequently the location of the target wearing one of these transmitters/receivers. A master node sends acoustic signals to the others to localize themselves with respect to the master node with the power and the direction of arrival (DOA) of signals received by the two microphones on the smartphone. An iterative expectation-maximization method is then used for the nodes to communicate their local estimate to the rest of the nodes. While the reported results appear to be excellent, they are produced under a completely static network—an assumption incompatible with most realistic scenarios. Static nodes are also used in the evaluation of the method outlined in [14], which utilizes a trilateration algorithm to localize each node in a multi-sensor network after converting the received signal strengths to meters via a path-loss model. An interesting feature of this algorithm is that it incorporates a means to temporally align the collected signal strength readings. In [15], an anchor-free self-calibration is proposed by means of an “Iterative Cole Alignment” algorithm based on a spring-mass simulation [16], and ensures synchronization by assuming that all receiving devices are linked to the same laptop; this method was evaluated assuming that the target always remains within a confined area.

The general question of how localization methods are evaluated arises in many publications, including the ones we discussed above. For example, static node configurations and artificially confined locations for the targets are fundamentally unrealistic and are bound to fail in the scenarios motivating our research. In this study, we collect sensor data resulting from the movement of one (or two) individual(s) in an actual apartment, following real (albeit scripted for the sake of consistency) movement scenarios throughout this apartment.

Indeed, when trying to use data collected from multiple sensors for location estimation (and activity recognition), a noticeable problem is sensor synchronization: most of the time, the clocks of the emitting sensors and devices involved are not completely synchronized. The approach proposed in [17] assumes that multiple sensors—each with its own clock and all connected to a single host—collect their timestamped observations in a FIFO (First-In-First-Out) structure; the host fetches the sensor data and performs a reconstruction of the sensor sample times, assuming a constant drift for each sensor and deterministic communication times. In our work, synchronization is not explicitly solved at the data-collection step; instead, we introduce the concept of a time “window” which abstracts the timestamp units at a coarser granularity and allows our method to ignore the imperfect synchronization of the data sources/sensors. As we will see, the window size can have a substantial effect on the accuracy of results.

The field of multiple person location estimation has received less attention from researchers. The majority of work in this area has been limited to counting how many occupants there are in the space. For example, [18] uses only binary sensors to count the people present within an area, eliminating the outliers due to incorrect sensor firing. The algorithm is initialized with the assumption that only the minimum number of sensors are outliers, and repeatedly increases the number of outliers until a solution is produced. Unfortunately, this method cannot recognize two occupants when their movement paths cross. [19] uses RFID tags and readers for the same person-counting task: the method maintains an uncertainty circle for each person, with a radius computed as the product of their speed of movement multiplied by the time lapsed; when a new event comes from a reader, the method assumes the person most likely to have moved in the vicinity of the reader based on their uncertainty circle and their direction of movement. A more recent paper by the same group [20] uses a much more expensive Kinect sensor to actually estimate the occupants’ locations.

When using motion sensors, it is important to know how to place them to achieve the best accuracy while minimizing the cost. [21] proposes a method for optimizing the placement of PIR sensors in order to meet the desired accuracy requirements while minimizing the cost. Their procedure hierarchically divides the space to sub-areas, based on walls and other obstacles such as large static furniture pieces. It then superimposes a grid on the space, whose length is determined by the accuracy needed, and solves the optimization problem of placing sensors so that the maximum possible number of grids cells are covered. In our group’s previous work, we developed a sensor-placement method that optimizes the information gain obtained by each additional PIR sensor placed in the space [1].

The dream of indoor location sensing has always been just that – a dream. The difficulty of Wi-Fi tracking and other technologies has made it hard for anyone – from businesses to regular users – to figure out where they were in a venue. But the folks at Estimote, a Polish beacon company, may have just cracked the code.

Using something they’re calling “Nearables,” as well as standard Beacon technology, the company is now able to track people and objects in a building. What does this mean?

First, you won’t be able to be tracked without your permission but once you’re on the network you’ll be able to see where you are on a map, identify objects anywhere in a venue, and even see where your friends and co-workers are anywhere in the building. Think of Harry Potter’s Marauder’s Map without the ability to track cats (unless you stick a nearable to them).

The system works by triangulating your device’s location via three or more beacons. You can try out the new indoor location app here and read a bit more about it here. Recently the company announced a partnership with Target that will bring indoor location to stores over the next few years.

From the blog post:

How do we know a nearable’s position, you ask? This is where the magic of Estimote Cloud kicks in: any time a user of the Indoor Location app enters range of a sticker with the app in the foreground (background mode coming), that nearable’s position is saved in the cloud. It works even if the nearable is private. In this case it won’t be visible for anyone except the owner, but everyone will be still be passively updating its location.

Adds Estimote co-founder Steve Cheney: “We built search for the physical world. You can literally search for objects tagged with nearables – they will be highlighted on a map with the relevant location, as long as the location itself is either public or belongs to you. This works like magic: any time a user of the Indoor Location app (in the future any app with our SDK) enters range of a sticker with the app its position is saved in the cloud. Even private nearables can still passively update their location,” he said.

“Our main message today is that Estimote is not a beacon company…We’re a full stack location intelligence platform and have successively brought together many different products – hardware, cloud software, device SDKs and data science – to create a developer-friendly platform for location intelligence and context.”

There are a number of competing systems to this including attempts by Apple and Google to improve indoor positioning. It’s a hard problem to solve. The goal, ultimately, is to hide beacon technology in almost everything.

As the technology gets smaller the company expects that the data points – and hence the location data – will be far more precise. By tracking items in a store and the shoppers you can see where people linger, where they rush by, and what exactly they’re thinking when they leave a cold pack of hot dogs in the cookie aisle. That last part, I suspect, will require more technology than humankind possesses.

0 thoughts on “Estimote Indoor Location Example Essay”


Leave a Comment

Your email address will not be published. Required fields are marked *