ION GNSS 2012
Session F5: Consumer Products & Services

Title: A Comparison of Feature Extraction and Selection Techniques for Activity Recognition using Low-cost Sensors on a Smartphone
Author(s): S. Saeedi, Z. Syed, N. El-Sheimy, University of Calgary, Canada
Date/Time: Friday, September 21, 2012, 9:20 a.m.
Room: 206 (NCC)

Recent popularity of mobile devices has resulted in considerable research efforts directed towards the recognition and monitoring of dynamic activity patterns using the low-cost sensors embedded on a mobile device such as smartphones and tablet personal computers (Tablet-PCs). This has been motivated by a number of important applications including context-aware mobile applications (such as location-based and navigation services), the health-care and security agencies. While different motion sensors on mobile devices (e.g. accelerometer, gyroscope, magnetometer, etc.) enable the collection of a vast amount of information about the user in an automatic way, it is still difficult to organize and aggregate all the collected information in a coherent, expressive and semantically-rich representation. In other words there is a gap between low-level sensor readings and their high-level context description such as the user´s activity. The activity contexts that are important in navigation applications include the user´s activity (e.g. walking, running, stationary, etc.), transit mode (e.g. driving a car, bus, train, etc.), and space (indoor or outdoor environments) while considering the various possible placement of the mobile device (e.g. in hand, on the belt, in bag, etc.) with an arbitrary orientation. This paper aims at finding the optimum set of sensors and features that contributes in an accurate and robust activity recognition methodology for a context-aware ubiquitous pedestrian navigation service for different user´s modes and motions.
Activity recognition systems which employ fusion of different sensors typically follow a hierarchical approach. The sensors and data providers collect and track useful data and information about user´s motions at lower levels. The next step is to extract features and characteristics of the raw measurements using statistical techniques. Finally, a classification or pattern recognition algorithm is used to recognize the user´s activity based on the comparison of the extracted features with those that are already extracted for each mode. In general, features can be defined as the main characteristics of a data segment that accurately represents the original data, while the feature extraction is the process to identify valid, useful and understandable patterns in raw data. If a good feature set is selected, a simple classification technique could efficiently yield accurate activity recognition; on the other hand, a poor feature set may need a complex non-linear classification technique whose structure is difficult or impossible to discern. There are no theoretical guidelines that suggest the appropriate feature set to be used for specific activity recognition situations. Therefore, the main contribution of this study is to provide an experimental guideline for finding the most meaningful features of motion sensor´s data and eliminating the redundant information.

In this research, to increase efficiency of the activity recognition and reduce computations, a comparison of different feature spaces and feature selection methods has been conducted. In the first step, various feature extraction techniques have been explored including Time-Domain, Frequency-Domain, Time and Frequency-Domain and heuristic features classification. After this step, large number of features are extracted, some of which not only provide irrelevant information for activity recognition but also increase the computation cost and training time. Therefore, feature selection algorithms have been used to find the optimum and independent set of features for each activity. The feature selection approach consists of detecting the features and information that are proven to minimally produce a correct response by the activity classification. The main feature selection algorithms in this study include principal component analysis (PCA), information gain, nearest neighbor (k-NN), and support vector machine (SVM). The ratio between inter-class and intra-class distances is then maximized across various feature subsets. Other criteria can be devised by analyzing the classifier output such as Bayesian network or multi-layer perceptron (MLP) artificial neural network (ANN).

Extensive experiments were conducted in different indoor and outdoor test scenarios to investigate the capabilities of different feature extraction and selection methods. The main sensor observations include assisted-GPS measurements, accelerometer, gyroscope, camera, temperature sensor, barometer, magnetometer and WiFi received signal strength (RSS). An android smartphone and a portable prototype device developed in Multi Mobile Sensor System research group at the University of Calgary were used in the data collection. In this study no assumptions were made about how users carry the device; so, the device was carried in different orientations such as in hand, in pocket, in bag or on belt. Various feature selection techniques were applied using adaptive window-sizes and different low-pass filters. Then various linear and nonlinear sensitivity analyses were conducted using dimensionality reduction. The results of each were used as an input to activity classification algorithms. Using feature selection methods, the set of twelve features were selected instead of possible 150 features with the same accuracy for activity recognition. Our results demonstrate considerable improvements in accuracy and time efficiency by the use of simple classification algorithm and using a small set of features and sensors. The results show that the relative WiFi RSS measurements integrated with temperature and barometer sensor have a good performance for indoor and outdoor context detection. Also accelerometer, gyroscope and GPS measurements can identify the mode of transportation. Additionally, accelerometer and gyroscope data integrated with magnetometer sensors result in activity recognition rate of almost 90% for different user´s activity and device orientation. The analysis and comparison of various feature extractions and selection methodologies provide useful insight into the activity recognition for smart location-based services.

Previous Abstract Return to Session F5 Next Abstract