MAMS: A Mobile Application to Detect Abnormal Patterns of Activity

1
MAMS A Mobile Application to Detect Abnormal
Patterns of Activity
Omar Abdul Baki Ying Zhang Martin Griss Hsiuping
Lin CyLab Mobility Research Center
Mobility Research Center Carnegie Mellon Silicon
Valley
2
Agenda

• Introduction to Anomaly Detection
• Related Work
• MAMS
• Experiments
• Results and Analysis
• Conclusion and Future Work

3
What is an anomalous incident? In the context
of anomaly detection, it is an activity which
isnt part of an individuals regular routine.
4
Anomalous Activity Detection in the Real World

• Detecting anomalous incidents amongst seniors in
  real time
• To notify caretakers sooner when
• something serious occurs.
• To reduce the stay-at-home costs
• for seniors.
• Securing Mobile Devices
• Providing an added level of security into Mobile
  Devices
• Making sense of changes in daily patterns of
  activity
• Uncovering changes in quality of life amongst
  seniors
• To detect the onset of certain medical conditions
  (i.e Alzheimer's)

5
Problems

• Most implementations use supervised learning
  approaches
• May not detect events outside the learned scope
• Some learned activities are user dependent and
  dont generalize well to other users.
• Enumerating and training a thorough set of
  learned activities is difficult and time
  consuming
• Such implementations are impractical since they
  would require an active end-user for training.
• Most implementations fail to account for location
  of events
• May generate false alerts for events which are
  location dependent

6
Related Work

• Unsupervised, Dynamic Identification of
  Physiological and Activity Context in Wearable
  Computing, Krause et. al., 2000.
• Towards Recognizing Abstract Activities An
  Unsupervised Approach, Hein A., Kirste T., 2008.
• Unsupervised Clustering of Free-Living Human
  Activities using Ambulatory Accelerometry,
  Nguyen et al., 2007.

7
MAMS Overview

• Developed on Nokia N95 platform using Mobile
  Python.
• MAMS uses 3 features to classify atomic
  activities.
• Location
• Based on REDPIN - indoor WIFI-based positioning
  system.
• Movement
• Variance in Accelerometer axes readings
• Posture
• Mean readings in Accelerometer axes readings

User Interface
Learning Logic/ Abnormality Detector
Application Logic
MAMS Cluster System
Sensor Sampler/Aggregator
WI-FI Sensor Interface
Accelerometer Sensor Interface
8
Clustering Algorithm

• Data points corresponding to the same atomic
  activities form clusters.
• New data points clustered based on a Euclidean
  distance measure.
• Clustering is continuous and incremental.

9
Experiments

• 1. Normal Event Log Collection
• 5 subjects to carry a mobile device for 3 days
  while performing daily routine.
• Subjects expected not to perform any irregular
  activities.
• 2. Abnormal Event Log Collection
• Subjects perform a predefined set of abnormal
  activities.
• 3. Compute the systems precision and recall when
  calibrated for optimum performance.

Number of labeled abnormal
activies correctly classified by MAMS Precision
Total number of labeled
abnormal activities classified by MAMS
Number of labeled abnormal
activies correctly classified by MAMS Recall
Total number of labeled
abnormal activities
10
Experiments
11
Results
12
Results
90 precision, 40 recall
13
Results
14
Conclusions

• MAMS is an anomalous activity detector based on a
  KNN unsupervised clustering algorithm
• MAMS supports continuous and incremental learning
• MAMS produces user-specific activity models
• MAMS distinguishes well between anomalous
  activities and normal ones

15
Future Work

• Analyze performances of several alternate online
  unsupervised learning methods
• Test MAMS performance in target environment.
• Re-valuate the performance of the location
  feature in larger experiment
• Evaluate the performance of the classifier with
  new features (ie. time of day)