Datenbestand vom 13. Juni 2019
Tel: 089 / 66060798
Mo - Fr, 9 - 12 Uhr
Fax: 089 / 66060799
DER VERLAG IST IN DER ZEIT VOM 12.06.2019 BIS 23.06.2019 AUSCHLIESSLICH PER EMAIL ERREICHBAR.
aktualisiert am 13. Juni 2019
978-3-8439-0113-0, Reihe Informatik
Ulf Mario Blanke
Recognizing Complex Human Activity Based on Activity Spotting
174 Seiten, Dissertation Technische Universität Darmstadt (2010), Softcover, B5
This thesis investigates methods for recognizing physical activities from wearable sensors. Human activity is a pivotal ingredient for context-aware systems. Systems can profit from context, in particular from the user’s activity by understanding what the user is currently doing and what goal he is aiming at.
In this thesis we focus on three main challenges for activity recognition. First, we explore methods for spotting sporadically appearing atomic activities from a continuous data stream - for example, drilling or screwing during a construction task. Large quantities of irrelevant data, the variability in execution between users and the diversity of activities render activity spotting a difficult task.
Second, we investigate the recognition of composite activities, that can be decomposed into several sub-activities as, for instance, fixing two parts is composed by drilling and screwing. Composite activities impose new challenges, as they usually contain more variability. Different durations in execution, interruptions or changing order are just a few characteristics of the variability.
Third, we address the reduction of training effort. One of the predominant problems is that activity recognition systems are trained from scratch for each single activity. Reusing acquired knowledge by transferring knowledge to related recognition tasks has been identified as promising method to minimize the learning effort and to use time-consuming annotations more efficiently.