1. WalkSense: Classifying Home Occupancy States Using Walkway Sensing
Elahe Soltanaghaei, Kamin Whitehouse
Department of computer science, University of Virginia
Thank you for the intro, and all of you for being here. As you all know motion sensors are the most common way of detecting occupancy in buildings today, in this talk I’m trying to convince you that we could be using them in a better way.

2. Motion ≠ Occupancy
Today the way that we deploy motion sensors is to put one sensor in every room. However, motion sensors are notoriously unreliable. They can only detect people if they are moving in the field of view of the sensor. However, they fail to detect a person’s presence who is out of field of view or sitting still which creates ambiguity about which zones are occupied
These issues happen because motion sensors are not acctually design to detect occupancy, but detect motion.

3. Motion ≠ Occupancy
Detecting occupancy in sleeping zones is even more difficult because people remain still for a long period of time while sleeping. It becomes even worse when there are multiple people at home. For example, we can detect people as long as they are moving around, but as they settle down, we might get motion events once every hour or so, which makes it ambigueues where they exactly are.

4. Walkway Sensing
So, we do not rely on motion sensors to detect occupancy inside rooms, but use them to detect motion in walkways between rooms such as doorways, foyers, or halways. We hypothesize that motion sensors work reliably in walkway for two reason: First, because people are always moving in walkways and do not sit still in them for long periods. Second, because walkways are small enough to be within the view of a motion sensor.
But, we still don’t know if they are going inside bedroom or coming out. So, we only need to install one motion sensor in the main active area to detect direction. The key point is that we need to detect the presence of people in active area only once duing their presence.
But the question is that how we can convert walkway sensing into a realiable form of occupancy sensing

5. Overview Active Sleep Away
The basic insight is that the occupancy states of the house can be deteceted through distinct occupancy zones: people are active in living areas, sleep in bedroom or leave the house when they are away. So, We use walkway sensing to detect when people move between these zones: so one walkway sensor between sleep and active zones, and another between away and active zones and to disambige which direction people are going, we put one more sensor in the active zone.
Scaling…example number of sensors. We hypothesise that walkway sensing not only detect occupancy with higher accuracy, but with fewer sensors.
Bedroom
Bathroom
Living room

6. Outline WalkSense: Occupancy detection algorithm
Automatic training feature
Evaluation

7. WalkSense Candidate Transition Events Active-zone events Sleep Active
Walkway
Active-zone
events
Time
WalkSense determines the current state of the home by tracking the sensor events in three zones and performs a real-time occupancy detection. In this example we have an active sensor and a sleep walkway sensor. But the system works similarly for detecting away periods. As default the home is in the active state,
So the system tracks the motion sensor events by moving along a state machine. if a person is detected in a walkway, system checks to see if it’s a candidate transition event, which means a walkway event with no activity in the other zones for a time window after the event. In this case, it’s not a candidate transition event and the system remains in active state and continue tracking sensor events. As a candidate transition evnet is detects a classification model will be executed to calculate the probability of a transition event into or out of the sleep or away state and as such it determines the current occupancy state.
The home remains in the sleep state until a sensor triggers, but it shouldn’t react aggresively to any sensor event as it might be only a false sensor firing or a short transition, such as person wakes up to drink water or go to the bathroom. Therefore, at this point the classifier will execute again to calculate the probablity of a state transition. And similarly, the sensor events will be tracked by moving along a state machine and define the current home status.
Active
Sleep

8. Training WalkSense How to train the models?
By exploiting the WalkSense in offline mode
However, as you know any supervised or clasification approach requires a training set wich is the labeled data. The conventional learning systems such as Nest thermostats assumes some sort of user involvement to report their sleep or away times for a period of time until the system learn the patterns, In addition, if the occupancy patterns of the home change it again needs to be reported by the users, Which makes it very user intrusive.
To solve this challenge, we exploit the walkSense algorithm in the offline mode and will use it as an automatic labeling or training mechanism. The intuition behind this idea is that the offline WalkSense has access to all the past data before, after, or during a sleep or away period and can determine the sleep and away periods with very high accuracies which makes it suitable for labeling. But first of all, lets see how it’s very accurate.

9. Training WalkSense
Candidate Sleep/Away Intervals: Inactivity longer than K
Candidate Away Intervals
Outside
Walkway
Active-zone
events
Time
Sleep
Walkway
In the offline mode, we try to find the home occupancy states in the historical data. So, we have access to all the past data in the active, sleep and outside zones. Then, we have defines a candidate sleep/away interval to be the duration between a pair of consecutive sensor events in a given walkway, which is longer than K minutes.
we de- fine the parameter K as the minimum duration of a sleep or away state and say that any sleep or away period shorter than K minutes can be safely ignored since the HVAC system and thermal properties of the home result in a certain time lag until the temperature setpoint is reached. As a result, the system fails to save energy and may even consume more by reacting to short away or sleep periods. Therefore, we expect reasonable values of K to be between minutes
That interval is decided to be in sleep/away state if no motion is detected in the other zone during this interval. Accordingly the sleep periods and the away periods will be determined.
This feature provides an automatic training with no user involvment. In addition, we can improve the classification models over time with an incremental training to learn any changes in the occupancy patterns of a house.
Automatic Labeling  No user involvement
Incremental Training  Improve detection model over time
< K
Candidate Sleep Intervals

10. Experimental Setup 350 days worth of data House B House A House C, D
Kitchen
bedroom
Living room
bathroom
Kitchen
bedroom
House B
Living room
Kitchen
bedroom
House A
Living room
Kitchen
bedroom
Aruba
Living room
House E
Kitchen
bedroom
Study room
Living room
Active Sensing Zone
Sleep Walkway Zone
Outside Walkway Zone

11. Experimental Setup Baseline: HMM-base Evaluation Metrics
Energy penalty: away or sleep periods are wrongly detected as Active
Comfort penalty: Active periods are wrongly detected as Sleep or Away
The HMM is depicted in Figure 4(b). The hidden variable (yt) is a distribution over the home state: Away, Active and Sleep and the HMM transitions to a new state every five minutes. The observed variables xt are a vector of three features of the sensor data: (i) the time of day at 4-hour granularity, (ii) the total number of sensor firings in the time interval dT, and (iii) binary features to indicate presence of front door, bedroom, bathroom, kitchen, and living room sensor firings in the time interval dT. The first feature helps the HMM use historical occupancy at each time of day to help estimate current occupancy. The second feature simply indicates whether the occupants are highly active. The third feature helps detect whether the occupants have opened or closed a door recently, and also helps filter out motion sensors with high false positives, e.g. those that are near a window.

12. Experimental Setup
61% of the day outside home, or sleep

13. 32% lower comfort penalty
Evaluation
30% lower energy penalty
32% lower comfort penalty

14. Evaluation 95% vs. 83% Accuracy WalkSense HMM Baseline Actual
Predicted
the Online-HMM approach misclassifies 10% of sleep and away events in place of each other
95% vs. 83% Accuracy

15. Analysis
Training Size Effect

16. Occupancy States & Occupancy zones.
Conclusion
WalkSense: Occupancy Detection algorithm based on Walkway Sensing
Higher Accuracy
Fewer Sensors
96% Accuracy with 30% and 32% lower energy and comfort penalty
Limitation: Assumes direct relationship between
Occupancy States & Occupancy zones. TV
Study Desk
Limitation: occupancy states corresponds directly to occupancy zones.
Sutdy desk in the bedroom
TV in the bedroom
We can adjust that if we need

17. Conclusion: Generalizable

18. Thank You!