Project Manager: Prof. dr.ir. F.E. van Vliet
Faculty of Electrical Engineering, Mathematics and Computer Science
Project website: DS2D
The D2S2 project aims at developing an algorithmic framework for operating large-scale distributed sensor systems.
The D2S2 project focuses on the on-line monitoring and reasoning about complex spatio-temporal phenomena in large networks of mobile entities. A key, innovative feature of the project is the development and use of an advanced miniaturized radar sensor that can operate under a wide range of “difficult” environmental conditions (smoke, fog, etc.) that cannot be handled by typical localization systems in operation today.
The fundamental research question this project addresses is: How can we exploit large-scale and mobile distributed sensor systems to reason about aggregate spatio-temporal phenomena?
- To this end this project plans to contribute in the following three fields of research:
The development of an algorithmic and theoretical framework for programming large-scale distributed sensor systems in an adaptive and volatile manner such that the system remains dependable despite its mobile and generally unpredictable nature. (Embedded Software group).
- The development of a miniaturized radar sensor that can operate under a wide range of “difficult” environmental conditions (smoke, fog, etc.). This line of research is firmly rooted in the experience of TNO (radar) and UTwente (circuit design).
- The design and implementation of localization and tracking algorithms using the radar sensor and dependability IDE outlined above. This line of research builds on the experience of both partners from Delft.
As a motivating example, we focus on the dynamics of human crowds. In a crowd, pedestrians both influence and are influenced by exchanging sensory signals (e.g., visual and vocal cues), and by imposing physical constrains (e.g., limiting the availability of space and introducing inter-person pressure). These interactions are known to result to complex spatial temporal phenomena, such as the formation of lanes when the crowd moves in opposing directions. Alarmingly, the collective motion of crowds can also prove hazardous. Examples of hazardous behavior include the emergence of pressure waves in parades and blockage near exits.
In the particular case of emergency evacuations, a strong bias is frequently observed towards some of the exits, while the rest are not used to their potential. That is due to of a variety of reasons. Foremost, people tend to trust in majority decisions and thus follow each other. Another reason is panic. Panic epidemically propagates along a crowd and hinders its ability of thinking critically. Due to the dangers of collective motion, researchers have made significant effort to model the behavior of crowds. Through various models, they focused on studying, simulating, and predicting hazardous behavior. The research however concerns almost exclusively off-line analysis of crowd dynamics and its optimization through re-designing the crowd’s environment. On the contrary, we aim to guide the motion of crowds by observing their behavior providing feedback in real-time. We aim to do so by the use of wireless ad-hoc networks.
Our main assumption is that the crowd is device-enabled, i.e., that each pedestrian, or a sufficient portion of pedestrians, carries a wireless-enabled device. As such, we can deduce the crowd mobility from that of the wireless network. The volatility and sheer size of a crowd, demands that we avoid the overhead of a centralized solution. To ensure scalability, we compute and store all information on the network itself in the form of scalar fields that describe a local quantity, i.e., information potentials. Due to the smoothness and modality properties of these fields, nodes can be directed towards appropriate exits by means of greedy distributed primitives that operate on the observable gradients.
Local reasoning in large-scale dynamic networks
Think of the information on a network as a landscape: hubs are peaks, arrows reveal the local information gradient. The information (value of each node) may be sensor data (e.g. temperature, vibrations, crowd panic) or relate to network status data (traffic, energy). In the image below, the landscape is shown through the colored Voronoi graph of the network. The color of each Voronoi cell reveals the value of the information of the node in its center (in this case the node degree). In that sense, the peaks (in blue) of the landscape are the areas of rich information, the valleys (in white) are less interesting, and we can find short paths towards peaks by following the path of steepest ascend.
We study how local rules reshape an information landscape, revealing fundamental information trends. I.e., we propose decentralized algorithms for changing the architecture of a landscape, essentially controlling the influence of various phenomena (peaks/valleys) according to their relative importance (scale).
In this video, the network reveals its most connected area by reshaping the node density landscape.
Using the information landscape, the nodes of a network can answer questions relating to global phenomena using only a local view.
- Search/routing -- “Can I find short paths towards the areas of rich information/higher values?”
- Reasoning/inference -- “My value is larger than that of my neighbors. Is this a local or a global phenomenon? ”
Project duration: 1-2-2012 / 1-2-2016
Project budget: 500 k-€ funding
Number of person/months: person months
Participants: TNO, TU Delft, UT
Project budget CTIT: 182.7 k-€ funding
Number of fte’s CTIT: 1.2 fte / year
Involved groups: Integrated Circuit Design
CTIT Research Centre: Centre for Array Technology