Methods and Software Architecture for Activity Recognition from Position Data: Applied to Cow Activity Recognition

Publikation: ForskningPh.d.-afhandling

Dokumenter

  • Torben Godsk
    Torben GodskDanmark
This thesis describes my studies on the subject of recognizing cow activities from satellite based position data.

The studies comprise methods and software architecture for activity recognition from position data, applied to cow activity recognition.

The development of methods and software architecture is the first step towards a decision support system for dairy farmers. The system is to include a component for recognition of cow activities. The decision support system will -- in its final version -- convert the result of the activity recognition to knowledge about the cows' individual behavior, health and welfare state, in order to help dairy farmers monitor their herds.

The methods extract the information from the position measurements and optionally from other sensor measurements as well. This information is translated into knowledge on, which activity was performed by the cow, during the recordings of the measurements. This is done by describing the movement taking place between two positions. These descriptions of the given movements are then assembled in segments of a given size. A series of statistical calculations is performed on the ensemble of movements residing within the segment. The results of these calculations are applied to a given standard machine learning algorithm, and the activity, performed by the cow as the measurements were recorded, is recognized.

The software architecture integrates these methods and ensures flexible activity recognition. For instance, it is flexible in relation to the use of different sensors modalities and/or within different domains. In addition, the methods and their integration with the software architecture ensures both robust and accurate activity recognition. Utilized, it enables me to classify the five activities robustly and with high success rates -- both in an offline setup as well as in real-time with continuously streamed data. In real-time, I am able to classify the five activities with success rates of: 97.8%, 85.2%, 84%, 93.7% and 72.2%, respectively, with a weighted average of 90.6%. Moreover, when doing the classification offline, I am able to classify the five activities with success rates of: 98.6%, 90.4%, 95.7%, 91% and 85.2%, respectively, with a weighted average of 94.8%. Furthermore, the software architecture integrates with "PerPos: a platform for pervasive positioning". PerPos is a software platform. Through cloud services, it provides functionality for pervasive positioning applications, including activity recognition.
OriginalsprogEngelsk
UdgiverDepartment of Computer Science, Aarhus University
Antal sider224
StatusUdgivet - 2011

Note vedr. afhandling

Supervisors: Kaj Grønbæk, Mikkel Baun Kjærgaard, Bo Eskerod Madsen, Flemming Skjøth, Thomas Andersen

Se relationer på Aarhus Universitet Citationsformater

Download-statistik

Ingen data tilgængelig

ID: 40377431