Sensor Networks is a burgeoning field of Computer Science in which small, wirelessly linked devices sense and reason in a collaborative fashion about the world around them. Our current research is on effective sensor fusion; integrating the readings from multiple sensors to provide better information than using each reading separately.
For each sensor type, we construct a sensor model, represented as a probability density function, by sampling over the binary sensors’ sensing region to derive a description of the sensor types’ triggering patterns. We then use these models, along with Particle Filtering, which is a technique for probabilistic reasoning, to perform fusion among many sensors and isolate a moving object.
We have applied this work to the tracking of an intruder in a room. In this problem, an intruder can enter, move around, and leave a room. We use the triggering patterns of X10 passive infrared sensors, which have very poor sensing attributes, along with our Particle Filter to locate and track the intruder. The X10 sensors have range of half of the room, are blind for eight seconds after triggering, and only provide a binary reading. We document the intruder’s progress by taking photos with a robotic webcam.
I am working on this project under Ken Goldberg in the Alpha Lab.
Related Material
Automated Intruder Tracking using Particle Filtering and a Network of Binary Motion Sensors. Jeremy Schiff and Ken Goldberg. SysLunch. UC Berkeley. October 2006. [2.8MB .ppt]
Automated Intruder Tracking using Particle Filtering and a Network of Binary Motion Sensors. Jeremy Schiff and Ken Goldberg. IEEE International Conference on Automation Science and Engineering (CASE). Shanghai, China. October 2006. [327KB .pdf] [2.4MB .ppt]
Class Project: CS262A - Advanced Topics in Computer Systems - LARE: An Architecture for Locating And Reacting to Entities with Heterogeneous Sensors and Actuators [1602KB .pdf]
|
This figure shows the general premise of our system - using sensor firings to track an intruder traveling around a room. Intuitively, the more overlap of triggering sensors, the more likely an object is at a location. Similarly, the more overlap with non-firing sensors, the less likely an object is at a location. |
|
|
This figure depicts a trace of locations where photos are taken. The small dotted line is the path of the intruder. The large dots are places where the robotic camera would take a photo. |
|
|
This figure is a sample of an actual image taken during one of our preliminary tests. It shows me grabbing an object and trying to exit the room without being seen. |
