Chapter 3 Design Guidelines for the Classification Environment
A typical recognition system as shown in Figure 3.1 has many components. A large amount of raw data is initially collected for the diﬀerent activities that are to be classified. Using all of this information for activity classification is costly, complex, and ineﬃcient. Certain fea-tures are extracted from the raw data that help best separate the activities being classified. Feature extraction of the raw data can be done through manual and latent (hidden) pro-cesses. Manual feature extraction is a means of preprocessing the data to reduce the amount of raw data being provided to the classifier. In this process statistics such as maximums and minimums are collected from the raw data. The manual feature extraction leads to the gen-eration of activity inputs which are later processed by the classifier. Depending on the type of classifier being used, a latent feature extraction is performed by the classifier to separate the activities and to develop a set of rules to diﬀerentiate amongst these activities. Once these rules are established unknown activities are processed by the classifier and assigned a class label denoting the activity they best represent .
When developing a system for a particular context application there are many design deci-sions. This paper is primarily concerned with the activity recognization aspect of the system, while Martin et al  discuss a complete design framework for wearable electronic textiles. As in  our goal is to generalize the design and perform minimal tuning for individuals. The SVD-based approach is an e-textile solution for which several garment sizes will be con-structed. All experiments for this thesis were conducted on a pair of e-textile pants which is shown in Figure 3.2.
Final System Design
Hardware design – The data is collected from an e-textile garment which is a pair of pants currently hosting eight sensor buttons which return a total of eighteen sensor readings. Ta-ble 3.1 provides a description of the sensor button statistics. The gyroscopes return angular velocity, the piezoelectrics return applied pressure on heel strike, and the accelerometers return the accelerations for the two axes for which they are aligned. While discussing the sensor orientation in Table 3.1 it is assumed that the X axis lies in front of the user i.e. it changes as the user moves forward or backwards, the Y axis lies along the user’s height, and the Z axis lies along the user’s side i.e. it changes as the user turns left or right. The data collected from each sensor button is relayed over an I 2C bus to a Bluetooth module which transmits the data to a laptop or PC. The data is stored in the form of ASCII text files which are parsed to generate MATLAB m-files. The MATLAB m-files represent the data collected for each activity as a list of arrays for each sensor button.
diﬀerent activities as well as a good aid for exploring the design space. Although most of the experimentation is done based on the data collected from the e-textile pants, the software model can classify activities obtained through alternate data files such as C3D files  and data collected using the Cubix . The classification environment generates inputs from the raw sensor data by pre-processing it over each channel of each sensor specified by the user. Figure 3.3 illustrates the approach taken to generate activity inputs for the classifier.
There are two important aspects involved in developing a good recognition system. The first is to obtain a set of desirable inputs by selecting the most appropriate sensing variables and the second is to generate rules from these inputs to obtain a high accuracy while classifying unknown inputs. The next section of this paper will examine the list of sensing variables that can be altered to generate activity inputs for the recognition system.
Factors Influencing Generation of Activity Inputs
Format of Data Files
As mentioned most of the experiments done to substantiate the results are performed using the e-textile pants. The classification environment can also use C3D files or data files from the Cubix for activity classification. The C3D files provide positional information for many body points and as described in  can be simulated to represent actual sensor readings. The e-textile pants were developed based on this simulation methodology. A large set of C3D files needed to classify the activities explored in this paper was not available at . Using the Cubix , activities of interest could be generated, but the Cubix could only handle four channels of information. Thus, most of the testing was done using the e-textile pants which could handle many more channels of information.
Classification Window Size
Clearly defining the objectives of the application is the first step towards developing a robust system. This helps in determining the required sensors for the application and the appropri-ate window size over which the system should collect sensor data. The classification window is the window over which data is captured for the purposes of activity input generation and classification. The window size should be as large as the activity or type of motion it is trying to recognize. For example, an elaborate dance may require a one to two minute window but a transition from the sitting to the standing position would require a very short window of less than a second to capture that specific motion. If the window size is five seconds and the sensors sampling rate is set at 120 samples per second then at least 600 samples from each sensor must be collected.
The location and type of sensors are chosen based on the application requirements. In , Van Laerhoven performed sensor placement testing using accelerometers placed on a Velcro strap and determined that the outside of the upper-leg, just above the knee, was the best position to collect data for basic everyday activities such as walking and running. Simulations could be run to determine the best combination of sensors and locations for a particular set of activities but that is beyond the scope of this thesis . Concentrating on just the number of sensors and not on the classification algorithm or robustness of the system, one could argue that including just the sensors that pick up the characteristic aspects of a context will be suﬃcient and adding sensors will not improve the recognition. Sensor fusion theory  on the other hand shows that recognition often becomes faster and more accurate as sensors get added.
1.3 Thesis Organization
2.1 Wearable Devices and Electronic-Textiles
2.2 Context Awareness
2.3 Activity Recognition Methods
2.4 Singular Value Decomposition
2.5 Principal Component Analysis
3 Design Guidelines for the Classification Environment
3.1 Final System Design
3.2 Factors Influencing Generation of Activity Inputs
3.3 Feature Extraction Algorithm
3.4 Factors Influencing the SVD Classifier
4 The Classifier – Vector Generation and Classification
4.1 Vector Generation Mode
4.2 Classification Mode
4.3 Real-time Classification
5.1 Selection of k Over Different Activity Sets
5.2 Change in Classification Error by Varying the Activity Inputs
5.3 Modifying the Feature Space to Improve Results
5.4 Longer Classification Windows and Activity Transitions
5.5 Final Online Results for 17 Activities
6.1 Future Work
GET THE COMPLETE PROJECT
Activity Recogniton using Singular Value Decomposition