Initial Goal

• Get several vehicles to autonomously converge on a point
• Very ambitious!!
  • Too many subproblems for one person in a few weeks
• Decided to scale back to just the localization subproblem
• In order to get good results, we would first need good localization

Modified Goal

• Default project with a small twist
  • Use Bluetooth at similar range to Wifi
  • Get readings simultaneously
  • Compare results
  • Try to combine if beneficial
• How will bluetooth perform at a range of several meters compared to WiFi?
• Can we use both to get a better result?

Data

• Bluetooth beacons placed to form smaller square than WiFi beacons to avoid interference from tables
• WiFi & BT RSSI measurements (100 samples each) taken at 32 points within the inner bounding box

Data - Raw

Data

• Clear variations in RSSI were visible
• Lots of noise
• Solution: Low Pass Filter
• Chebyshev Type 1 Filter chosen for steep roll-off

Filtered Data

Localization

• We have some pretty clean data to work with (in most cases)
• First: Channel model
  • Used the common channel model
  • rssi = -10n * log(distance) + power
  • Constants chosen experimentally to minimize error

def rssiDistance(rssi, power, n):
    return (10.0 ** ((power - rssi) / (10.0 * n))) * 100.

def blueDistance(rssi, power=-62.0):
    return 0.5 * rssiDistance(rssi, power, 2.6)

def wifiDistance(rssi, power=-32.0):
    return math.sqrt(abs(((240 ** 2) - ((0.8 * rssiDistance(rssi, power, 3.0)) ** 2))))

Localization

• Localization Algorithms (Tarrío, Bernardos, Casar):
  • Unweighted Hyperbolic
  • Unweighted Circular
  • Weighted Circular
• Methods:
  • Used numpy & scipy for computation and optimization
• Results:
  • Weighting was absolutely necessary
  • Best error function:

e = lambda i: ((distance(guess, anchorPositions[i]) - anchorDistances[i][-1]) ** 2) / weight[i]

Discussion

• Best Results (Weighted Circular):
  • Mean Error (WiFi): 1.4M
  • Mean Error (BT): 0.7M
• Unweighted Algorithms were not viable at all
  • In a space this small, not much variation in RSSI
  • Noise is a better indicator of distances of this order

Future Work

• WiFi worked best in the middle of the room (larger range)
• Bluetooth worked best closest to the beacons
• Try combining WiFi & Bluetooth
  • Weight Bluetooth very high if readings indicate close to a beacon
  • Weight WiFi very high if readings indicate middle of room
  • Should be easy to implement (change weight term)
• Try to implement the consensus algorithm!
  • Better filtering will be needed for moving nodes
  • Need to improve sampling rate
    • Write custom sampling code in C

Challenges

• Initial goal was too ambitious
• Took a lot of tweaking to get decent results
  • Was very discouraging for a while
• Working alone: horrible idea
  • I don't know much about signal processing
• Taking data was extremely tedious
• Was worth it in the end though

Questions?