ULTech to Observe Elderly’s Behavior Changes over Time in SH
133
To build the routine behavior model, we follow the following steps:
1. We first use training data collected during the previous period which was
treated as a baseline behavior period. We then follow an unsupervised learning approach: clustering to find point anomalies. To address this, we cluster
instances of each activity based on start time and duration without considering the day of the week. For clustering, we use the DBSCAN algorithm [11] which is a density based clustering algorithm. The major advantage of DBSCAN, compared with other clustering algorithms like K-means,
is that we do not need to specify how many clusters should be identified.
After clustering, DBSCAN marks each point as belonging to a cluster or as
noise(anomaly).
2. We eliminate point anomaly and we calculate the average start time and
duration for each activity in training data.
The activity “going to the toilets” that occurs several times a day is treated
separately. There are usually regular schedules for this activity. It is not essential
to be very precise on the realization time of this activity. We then choose to study
its frequency rather than the occurring time. Using the same training data to
model regularities in the three studied activities, we calculate the frequency, per
n hours, of the activity “going to the toilets”.
3.3 Elderly’s Behavior Change Detection
Once computed, the normal behavior pattern can be used to detect anomalies
by comparing the current behavioral data of an elderly with her/his normal
behavior pattern. The basic idea of our behavioral deviation detection system
is to estimate the similarity between both patterns using a score. We therefore
consider three criteria for the activities: the time, duration and chronological
order of the activities in the sequence.
Behavior Modeling Using a Daily Activity Score: Intuitively, a particular
activity is similar to a pattern if its start time, duration and location are similar
to the ones defined by the pattern. The similarity of the time and duration for
each activity is estimated by a score.
The similarity score of an activity a in a day d is calculated by the formula
(1). It is given as a percentage and represents the temporal intersection of the
normal behavior pattern and one observed day pattern, for the same activity.
We note that S ad is the start time of activity a d , D ad is the duration of activity
a d and E ad = S ad + D ad is the end time of activity a d
Similarity score =
(inf(E an , E ad ), sup(S an , S ad )) ∗ 100
D an
(1)
The similarity score for one day is the average of similarity scores for all
activities occurring in this day.
133
To build the routine behavior model, we follow the following steps:
1. We first use training data collected during the previous period which was
treated as a baseline behavior period. We then follow an unsupervised learning approach: clustering to find point anomalies. To address this, we cluster
instances of each activity based on start time and duration without considering the day of the week. For clustering, we use the DBSCAN algorithm [11] which is a density based clustering algorithm. The major advantage of DBSCAN, compared with other clustering algorithms like K-means,
is that we do not need to specify how many clusters should be identified.
After clustering, DBSCAN marks each point as belonging to a cluster or as
noise(anomaly).
2. We eliminate point anomaly and we calculate the average start time and
duration for each activity in training data.
The activity “going to the toilets” that occurs several times a day is treated
separately. There are usually regular schedules for this activity. It is not essential
to be very precise on the realization time of this activity. We then choose to study
its frequency rather than the occurring time. Using the same training data to
model regularities in the three studied activities, we calculate the frequency, per
n hours, of the activity “going to the toilets”.
3.3 Elderly’s Behavior Change Detection
Once computed, the normal behavior pattern can be used to detect anomalies
by comparing the current behavioral data of an elderly with her/his normal
behavior pattern. The basic idea of our behavioral deviation detection system
is to estimate the similarity between both patterns using a score. We therefore
consider three criteria for the activities: the time, duration and chronological
order of the activities in the sequence.
Behavior Modeling Using a Daily Activity Score: Intuitively, a particular
activity is similar to a pattern if its start time, duration and location are similar
to the ones defined by the pattern. The similarity of the time and duration for
each activity is estimated by a score.
The similarity score of an activity a in a day d is calculated by the formula
(1). It is given as a percentage and represents the temporal intersection of the
normal behavior pattern and one observed day pattern, for the same activity.
We note that S ad is the start time of activity a d , D ad is the duration of activity
a d and E ad = S ad + D ad is the end time of activity a d
Similarity score =
(inf(E an , E ad ), sup(S an , S ad )) ∗ 100
D an
(1)
The similarity score for one day is the average of similarity scores for all
activities occurring in this day.