Turning raw SenseCam accelerometer data into meaningful user activities
Qiu, Zhengwei and Doherty, Aiden R. and Gurrin, Cathal and Smeaton, Alan F. (2010) Turning raw SenseCam accelerometer data into meaningful user activities. In: SenseCam 2010 - second annual SenseCam symposium, 16-17 September 2010, Dublin, Ireland. ISBN 1872-327-915
Full text available as:
The onboard accelerometer is one of the most important sensors in the SenseCam, where it can influence the quality of photos captured by choosing the optional time to take pictures. Compared with other sensors, there are a number of advantages that the accelerometer has:
Acceleration data is easy to be stored and processed, especially in comparison to the average of 4,000 images taken every day by the SenseCam which consume an amount of disk space. Acceleration data takes little space, and can be processed by the SenseCam’s on-board micro-processer in real-time.
The most important information from SenseCam is image data, but it is difficult to take a clear photo when the user is moving very fast or is present in dark places, the accelerometer helps to avoid the problem of blurred image capture.
No wireless signals are needed. While GPS techniques have been improved a lot in the past decade, determining location of the inside of buildings is limited by no clear line of sight to satellites in the sky.
Low battery consumption. Compared with camera and GPS, it uses little battery. Now the battery is the key bottleneck for portable sensors. It is very inconvenient for users to constantly remember about charging battery all the time.
The accelerometer can also act as an important source of evidence for automated content annotation, as described below.
Given the above benefits of the accelerometer onboard the SenseCam, we now discuss the information which can be mined by analysing this raw acceleration data:
1. Activities detection: By analysing acceleration data, common daily activities can be recognised, such like sitting, walking, driving and lying. To recognise each different activity, we employ different binary-class SVM models for each activity, because different features from acceleration are used to recognise different activities. These activity recognition results can be used as an important resource for associating context with real-time lifelog information.
2. Calculate driving-related CO2: Environmental issues have been at the forefront of the public conscience of late. We believe it will be helpful if users can get real-time driving information and how much carbon they have produced by driving. To achieve this we have built SVM classifiers to identify driving related activity. Our driving detection can be improved by smoothing algorithms and also techniques to detect time spent at traffic lights. From this we can make an accurate estimation of driving related CO2 emissions.
Archive Staff Only: edit this record