Doorjamb - Network and Systems Lab

Report
Doorjamb: Unobtrusive Room-level
Tracking of People in Homes
using Doorway Sensors
Timothy W. Hnat, Erin Griffiths, Ray Dawson,
Kamin Whitehouse
U of Virginia
Presenter:SY
About This Paper
• Unobtrusive room-level tracking
– People in homes
• Doorway sensors
– Ultrasound sensor
• Method
– Estimates the height and direction
Technical Problems
• Multi-target tracking
– Data association
• Noise
– Person’s posture, multipath reflections, and the
natural undulation of gait
• Algorithms
– Crossing event detection
– Tracking
Contributions
• Hardware
– Design and prototyping
– Lesson learned
• In-depth analysis of the sources errors
– Present signal processing algorithm
• Data association challenges
– Tracking algorithm
• Proof-of-concept implementation, deployment,
and empirical evaluation
Outline
•
•
•
•
•
Hardware design
Signal processing algorithm
Tracking algorithm
Evaluation
Conclusion
Hardware
• Features
– Cost effective
– Battery powered
– Wireless
• Design
– Detect height
• Measure the distance to the top of the head
– Detect walking direction
• Angled into one room more than the other
Doorway Sensor
•
•
•
•
•
Parallax PING ultrasonic range finders
Passive infrared sensors
Magnetic reed sensors
Custom-designed power module
Synapse Wireless SnapPY RF100 module
Achieving Doorway Coverage
• Requirements
– 1 cm resolution
– Heights ranging from 151 cm to 189 cm
– Walking speeds up to 3 m/s^2
– Doorways range: 90-300 cm wide, 213-275 cm tall
• Parallax PING ultrasonic
– 40 degree beam angle
– Min: 2 cm; Max: 300 cm
Achieving Doorway Coverage
• Tallest person
– Gap between the head and doorway  24cm
– 40 degree beam  Sensing diameter of 17 cm
– Speed of 3 m/s, a head that is 15 cm diameter
• Pass sensing region in about 100 ms
– 50 Hz sample rate – one module at a time
Doorway size
• Typical doorway width of 90 cm
– Sensing diameter – 17 cm
– Head radius – 7 cm
– Two sensors should be enough
• Higher door frames require fewer sensor
• 300 x 275 cm
– 4 range finders
– Sampling rate 12.5 Hz
– Cannot support wide and short
Early Prototypes and Lessons Learned
Audible
click
Outline
•
•
•
•
•
Hardware design
Signal processing algorithm
Tracking algorithm
Evaluation
Conclusion
Signal Processing
• Input: stream of height value
• Output: doorway events D (tj,hj, vj)
• Four algorithms
– Doorway crossing detection
– Noise filtering
– Height estimation
– Direction estimation
Signal Captured
Doorway Crossing Event
• Find timeout, multi-path, measurement events
• Within 400 msecs of each other
Noise Filtering
Define clusters
Extend 200ms
Noise Filtering -- Obstacle
• Extends 30 seconds on either side
– Remove any height measurement that is positive and identical
Height Estimation
• Multi-path reflections
– Maximum measurement may fail
– Typically only occur once
• Height estimation
– If maximum height cluster exist
• Max of the cluster
– Else
• Maximum height
Direction Estimation
• Sensor tilts into the doorway
• Three algorithms
– Line slope
– Compare max height timestamp to median
– Compare min height timestamp to median
• Vote
– Each algorithm estimate: +1, -1, 0
– Sum all: [-3,3]
Outline
•
•
•
•
•
Hardware design
Signal processing algorithm
Tracking algorithm
Evaluation
Conclusion
Tracking
• Input: sequences of detection events D
• Output: Corresponding room states S, (r1i, r2i)
• Ambiguity
– False detections, miss detections
• Key insight
– Ambiguities can often be resolved by future
observations
MHT Algorithm
• Multiple hypothesis tracking approach
– Multiple alternative tracks are considered
simultaneously
• As new events are processed
– Tracks that are not consistent with the new
information are evicted
Overview
• Initial
– All tracks created with identical weight
– For 2 persons + K rooms, K2 tracks are created
• Update
– For each doorway event
• Update track
• Update weight (based on prior training study)
• Merging and Evicting
– Evicting low weight tracks
– Merging duplicate tracks
Prior Training Study
• Find conditional probabilities
– p(H|O) – a height measurement given the origin
– p(V|O) – a direction measurement given the origin
– p(H = ) – probability of missed detection
• Origin -- Person A, or B, or false detection
• Training period
– Each individual walks under each doorway
multiple times
Creating Tracks
• Initial tracks
– every possible combination
• For each new doorway event
– Between rooms i and j
– Five new states are possible
• a/b move to room i/j + false detection
– Duplicate every track 5 times
Weighting Tracks
• New weight is
– Old weight multiply by
– Probability of the origin moved through doorway
m given height measurement
– Probability of moving from room p to m given the
direction measurement
– Probability of moving from the last observed room
m-1 to p without having detected
Merging and Evicting Hypotheses
3
• “N-best” eviction
policy
– Keep the n best
tracks
• Problem –
duplicate tracks
• Track merging
algorithm
4
1
2
Outline
•
•
•
•
•
Hardware design
Signal processing algorithm
Tracking algorithm
Evaluation
Conclusion
Experimental Setup
• Built 43 ultrasonic doorway sensors
– Deployed across 4 different homes
– Periods of 6-18 months
– Used for development, testing, and iterative
design
• For this evaluation
–
–
–
–
–
Performed 3 controlled experiments
3 different pairs of testers
Randomly walk around
Collect ground truth with handheld device
3000 unique doorway events
Evaluation Metric
•
•
•
•
Type 1: correct state
Type 2: wrong person
Type 3: false room transition
Type 4: missed room transition
Tracking Accuracy
False Detections and Missed Detections
• Precision:
– The number of false detections divided by the number of total
detections
• Recall
– Number of missed detections divided by the number of true
doorway crossing events
Height Measurement Accuracy
Direction Measurement Accuracy
Systems Performance
• Average 24 states, max 55 states per track
• Real time, online
– With 500 ms look-ahead window
Limitations
• Fall short of true in-situ experiments
– Controlled experiments
• Do not capture long-term effects
• A proof-of-concept for Doorjamb tracking
• Scalability
– Typical homes with 3-4 people
• Requires calibration and training
• Does not detect children
Conclusion
• Track people in homes with room-level
accuracy
• Unobtrusive
• Achieve 90% tracking accuracy
• My opinions
– Well written complete work
– Not so sexy
– Has it’s own selling points

similar documents