Tracking Assignment

The goal of this assignment is to track people in a home using anonymous sensors such as motion detectors, break beams (door), contact switches, and pressure mats. We will provide simulated tracking data, and you need to provide us with sensor and dynamic models, and where people were and what state they were in (you define states).

The House

This house has 10 rooms (large numbers in diagram). There are 13 doors, monitored by break beams that sense movement direction. Movement from the low numbered room to the high numbered room is given by the door number (small numbers in diagram), and the reverse direction is given by the door number plus one. Each room has a drawer or device monitored by a contact switch (reads 0 if closed, 1 if open). Each room has a chair monitored by a pressure mat (reads 0 if empty, 1 if used).

Example of data format
This file is the raw sensor values triggered at each second by occupants. Each row corresponds to a second. The first ten numbers are the motion detectors for each of the ten rooms. The second ten numbers are the contact switches for each of the ten rooms. The next ten numbers are the pressure mats (again, one per room). The final 26 numbers are directional break beam sensors (2 directions for each of 13 doors).

Here is the C program I used to visualize the data. It uses OpenGL and GLUT. To compile it on LINUX type:
gcc simple.c -lglut -lGL -lGLU -lX11 -lm -L/usr/X11R6/lib -o simple
For Windows, here is a Visual C++ 6 project that contains the above program. GLUT for Windows is available from here.

Question 1: Estimate sensor and dynamic models, and tell us where people are, and what state they are in (you define states) for one person: one person data
Here is some more data to help you estimate sensor and dynamic models, if you need it: more one person data

Question 2a: (Easy version, YOU CAN DO THIS INSTEAD OF THE ORIGINAL QUESTION 2 (now 2b) OR BOTH) Estimate sensor and dynamic models, and tell us where people are, and what state they are in (you define states) for two people: two person easy data
We also provide each person alone with the SAME transition and sensor models used in the above data file with them together:
person A alone
person B alone

Question 2b: (Hard version) Here we don't give you individual data, and the transition models might not be all that different. Estimate sensor and dynamic models, and tell us where people are, and what state they are in (you define states) for two people: two person data
Here is some more data to help you estimate sensor and dynamic models, if you need it: more two person data

You ask: So what do I turn in?

Turn this in by sending a URL in email to cga@cmu.edu, sonyaa@cs.cmu.edu, neill@cs.cmu.edu and tingliu@cs.cmu.edu.
Provide a writeup that describes what you did, and why you did it.
Provide your software in a form we can run.
For each question (1, 2a OR 2b):
- At each second in the hour long data files, where was each person (room number)? If you had additional state, please identify the full state at each second in a separate file. For example, for problem 1 the location data should look like:
  9
  9
  8
  9
  9
  8
  8
  8
  8
  8
  8
  ...
- Dynamic model: This is the transition matrix, giving probabilities of state transitions p(next-state|current-state)
- Sensor model: These are the probabilities p(measurement|state). This is a giant table. Let's assume the sensors are local, so the probability a sensor fires when someone is not in the room is independent of where any people are in the house. This is the "false alarm" probability for each sensor. The rest of the sensor model is the probability a sensor in room X fires when the person is in room X. One minus this probability is the "miss" probability. Note that the miss probability should depend on whatever additional state you define, as well as the person's location.
Handling the doorways requires a little thinking, since that measurement does not depend on a "state" but a "state transition". How do you handle that?