It may feel like you are always on the grid and your location is constantly broadcast by some GPS device or another. In reality, we are often in environments, including buildings and parking lots, where GPS works poorly or not at all. That is why the possibility of analyzing the behavior of people and things in these environments captured the attention of experts from diverse fields.

Behavioral analytics was originally conceived to extrapolate online users trends (e.g., navigation paths and clicks) and to tailor marketing efforts accordingly. More recently, there has been a growing interest in systems that can analyze the behavior of people and things in public spaces, through dedicated infrastructures or by relying on people’s personal devices. Needless to say, the related privacy issues, the prerequisite for personal device availability, and the need for dedicated infrastructures make the design and operation of such systems challenging. Furthermore, target identification is not even needed for many applications. For example, business intelligence often requires monitoring the flow of people to figure out which exhibits at a museum are most popular, which areas of a building get the most use, or which part of a store receives the most traffic

For these scenarios where we do not need to find a specific person at a specific time, passive tracking is more economical and offers more privacy than active tracking. The figure below illustrates the progression of implementation cost savings and privacy preservation from active (e.g., cellular and wireless sensor networks) to passive (e.g., RFID and sensor radar networks) tracking systems.Figure 1 Active and Passive Tracking systems

Enter PATH

My work focused on developing a highly accurate and cost-effective way to track passive sources and passive targets for applications such as crowd counting and flow monitoring. Most recently, as part of a three-year Marie-Skłodowska Curie Fellowship from the European Commission,I created a project called PATH (PAssive Tracking of people and things for physical beHavior analysis) in collaboration with the Wireless Information and Network Sciences Group at the Massachusetts Institute of Technology, led by Professor Moe Win, and the Wireless Communication & Localization Networks Group at the University of Ferrara, Italy, led by Professor Andrea Conti.

The main goal of PATH is to define a new paradigm for the behavior analysis of people and things, by tracking their positions and dynamics with minimal implementation cost and maximum privacy preservation. Without relying on any personal device, PATH enables the detection, localization, and tracking of targets that do not necessarily participate in the localization process and integrates these capabilities in infrastructures for the IoT. This is useful for crowd-based decision making to enable the appropriate allocation of resources at critical times.

Building and Validating the Model

The main research activities carried out within PATH focus on deriving a framework for design and analysis of advanced techniques for detection and tracking of people and things based on reflections from the targets and on exploiting signals of opportunity (signals already on air for other purposes).

Designing and characterizing device-free localization in cluttered environments.

During the outgoing phase in the Wireless Information and Network Sciences Group within the Laboratory for Information and Decision Systems (LIDS) at MIT, I developed a general framework for the design of device-free localization systems which accounted for multi-path, clutter, and interference involving different targets. Based on this framework, I am currently deriving fundamental limits to determine the maximum localization accuracy that is achievable for given technology and environment. This is essential to analyze and design device-free systems for behavior analysis, e.g. to quantify a performance gap-to-be-filled and to design tracking algorithms that fill the gap.

Proposing low-level features for crowd-centric behavior analysis.

Current algorithms for physical behavior analysis rely on multi-target tracking. In multi-target tracking, a different set of measurements is associated with each detected target to estimate each target’s position and trajectory and behavior analytics are then extracted. This method is also known as individual-centric. However, individual-centric methods have high complexity that grows exponentially with the number of targets due to data association. This is unnecessary when the system is interested in flow monitoring and not individual target locations. For these reasons, there is a growing interest in designing crowd-centric methods for behavior analysis, i.e., methods that infer the number of targets and their behavior directly from the measured data without estimating their locations. In PATH, I developed a mathematical framework for the design of crowd-centric techniques for behavior analysis (see below).Figure 2 Example Scheme of Individual and Crowd-Centric Counting

In the first part of the project, the case of device-free counting was addressed. Counting targets in a monitored area enables several applications in a variety of scenarios including smart buildings, intelligent transportation, and public safety.  The counting problem has been addressed for both active and passive scenarios. In the active case, ultra-wideband (UWB) signals have been considered since they allow high spatial resolution.  For the passive case, OFDM signals-of-opportunity have been considered since they are available in most of the communication standards (DVB, DAB, WLAN, 4G, and soon in 5G) and they allow low complexity signal processing.

Introducing soft information (SI) algorithms for localization and tracking

The quality of the target position estimate depends on signal processing techniques and intrinsic network properties, including the wireless medium characteristics. While conventional localization and navigation approaches rely on estimating single-value metrics, higher levels of accuracy can be obtained by using soft information (SI), the ensemble of positional and environmental information, respectively, associated with measurements and contextual data. The SI can be extracted via sensing measurements (e.g., using radio, optical, and inertial signals) and contextual data (e.g., using digital map, dynamic model, and node profile). In PATH,a framework for the design and analysis of systems relying on SI was developed.  Such a framework was used to create a machine learning algorithm for crowd-centric counting and behavior analysis.

Validating the framework and the main algorithms through experimentation

The application of PATH’s theoretical findings use cases has been pursued throughout the entire project. The clutter removal and the machine learning algorithms have been validated through experimentation in an indoor office environment with UWB sensor radars. Results show that the techniques developed in PATH outperform current state-of-the-art techniques.

The main findings of PATH represent strong innovation in the existing state-of-the-art. Introducing a crowd-centric approach is a breakthrough for physical behavior analysis.

Visit the PATH project public website.