Project Details
Projekt Print View

Methods for Activity Spotting With On-Body Sensors

Subject Area Image and Language Processing, Computer Graphics and Visualisation, Human Computer Interaction, Ubiquitous and Wearable Computing
Term from 2009 to 2014
Project identifier Deutsche Forschungsgemeinschaft (DFG) - Project number 101885733
 
Final Report Year 2014

Final Report Abstract

Today, commercial wearable devices such as wristbands or belt clips can detect simple, repetitive activities such as walking or running. Extending that capability towards more complex activities (e.g., nutrition, industrial maintenance) is an active research area. This project was devoted to a particularly difficult aspect of this problem: the spotting of subtle actions within a continuous signal stream dominated by non relevant “NULL” class events. The difficulty stems from the variability of human actions: both the ones that we want to spot and the non-relevant NULL class ones which can encompass virtually any conceivable human activity. To address the above problem, the project has proceeded in three directions. First, we have investigated the value of different additional sensing modalities beyond motion sensors. In particular wrist worn cameras combined with proximity sensors (as can be easily realized in a smart watch) have proven to a be valuable source of information. In addition we have shown that a combination of head motion pattern and eye blink analysis using sensor integrated in Google Glass and a capacitive neck band can be used to reliably recognize high level activities. Related to direct sensor work we have shown that geometric body model which is derived from body-placed sensors allows to improve recognition results and is beneficial for recognition of fine-grained activities. Second, we have developed new spotting methods based on sequences of “basic motions”. This included novel signal segmentation methods based on a fast polynomial approximation, a concrete approach to identifying characteristic, user and situation invariant parts of the motion signal based on physical constraints or abstract “Eigenmotifs” and new approaches to classifying sequences of such basic motions. Third, we have shown that decomposing complex high-level activities into simpler actions, allows for transfer and sharing between these high-level activities and subsequently improves performance. To exploit fine-grained location information in the absence of location providing signals, we have shown how to estimate location based on detection of routinely visited locations and a pocket-based inertial measurement sensor. We also looked into multi-modal approaches with language. More specifically we generated natural language descriptions for activity sequences and we investigated different approaches and features to understand the similarity of activities by combining different state-of-the-art visual and textual features. Finally, we have developed an abstract characterization of different types of activity spotting problems and a new, more versatile way to evaluate their performance. Several data sets that were recorded during the project are being released into the public domain for other groups to test their methods on them.

Publications

 
 

Additional Information

Textvergrößerung und Kontrastanpassung