Total ingresos adicionales
1- SERVICIO CIVIL PUBLICO (LOSEP)
To fuse all sensor values, it is necessary to implement a function which can return the correct value at any point in time. The data is collected mostly from physical sensors and is not human-readable. In addition, it has different frequencies and generates a time gap. For example, available WiFi is scanned every two minutes. Physical sensors cannot return corresponding value immediately. WiFi can only generate one reading when the hotspot is found. One scanning action may take more than one minute. Accelerometers can generate more than five readings every second. But they do not work in real-time. There is always a time gap between readings. However, in some situations, the sensor value can be predicted. For example, a user’s current location can be predicted using his previous locations and location patterns. Because the software gathers different sensor data at different frequencies, the data is returned with different timestamps. Furthermore, raw sensor data is not easily shared. If a virtual sensor supplies the user with raw sensor data, all applications using it will also need to implement the data mining module. That will not only waste time, but it is also difficult to ensure that all users’ output is consistent.
In this study, we use five virtual sensors to solve above issues. They are time, location, activity, social and environment virtual sensors. As described in Chapter 2, a virtual sensor is a software sensor which has three functions: 1) combining different format data collected from multiple physical sensors; 2) transforming the combined different format data into a natural language which could be easily understood by humans; and 3) sharing the data through a web interface. To improve
110
the speed of the virtual sensors all sensors data are pre-processed. Virtual sensors can generate the meaningful semantic knowledge to better fit with a user’s expectations.
Time virtual sensor: The time sensor will return a different format of time depending on the user’s input parameters. For example, it can return “Morning”, when the user queries it with specific parameters. In this study, the time sensor is implemented by a segment code. It does not contain any data.
Location virtual sensor: The main function of the location sensor will supply the most accurate location to the user. Furthermore, the process is transparent to the user application. The output of the location sensor not only includes the address, but also other information such as home or work place. As mentioned above, the location information is calculated not only using GPS information, but also using WiFi and Base Station. For the location sensor, if GPS is available, the latitude, longitude and accuracy data will be returned to the user. The address and length of stay will also be returned. When the GPS is not available, the WiFi access point or Bluetooth will be transmitted to the GPS point. If the location information comes from a WiFi access point, the minimal time span will also be returned in case the user uses it to filter out some values. If GPS, WiFi and Bluetooth are not available, the Base Station will be considered as location information.
Activity virtual sensor: As mentioned above, accelerometers can be classified into four different activities as sitting/standing, lying, walking and driving. For the activity sensor, when a new accelerometer sensor data is read to the database, the SVM classifier will be executed and the results will be stored in the database. By requesting the activity sensor, the user’s activities such as
111
walking, sitting can be returned. For other activities such as phone calls, a message can be requested using a specific time range. For this sensor, user ID, time, time period and type are input parameters. The sensor will return different activities according to different type parameters.
Social virtual sensor: Once a new photograph is received, the system will judge if the photograph is blank. If the photograph is not blank, it will be processed by a face detection process. If a face is detected from the photograph, the number of faces and corresponding face’s position and size will also be stored. Bluetooth information was considered as an important social context data source. Almost every smartphone has a Bluetooth adapter. Bluetooth can be used to predict the relationships. Some users like to turn it off to save battery life. In our study, we found that 10% of users had Bluetooth switched on. During the pre-processing stage, the Bluetooth and location information were gathered for retrieval. The relationship between the device owner and user can be predicted. For example, if the devices appear in the user’s home and also in other place the owner of the Bluetooth can be the user himself or his friend. Social virtual sensor can be used by inputting user ID, time, time range, type. The face, relationship and related information can then be sent back.
Environment virtual sensor: The environment virtual sensor mainly supplies noise information such as “noisy” and “quiet”, although it can also acquire location information from the location sensor. With time and location information, it can locate the weather information and return it to the user. To enhance the response speed, all the weather information was downloaded and stored in the database when any new location was detected.
112
4.4 Summary
Physical sensor data is easily collected but very difficult to semantically understand as most of it is made up of numbers. It needs to be converted into semantic contexts such as personal, time, location, activity, social and environment contexts. This chapter introduced how these contexts were detected. In particular, we applied term weighting into the location context detection and the SVM into the activity context detection. Lastly, this chapter presented how to implement the virtual sensor in order to combine different format data; transform the combined data into a natural language; and to share the data through the web.
113
CHAPTER FIVE
SEGMENTING EVENTS
5.1 Introduction
This chapter provides a detailed explanation of the approach used to segment sensor data into events. It shows what attributes are extracted from sensor data based on which events are segmented. The experimental process and results are presented for testing hypothesis 1 which proposes that event segmentation can be performed effectively by detecting changes in sensor data. Finally, to select a suitable keyframe for each event, three users’ keyframe selection methods for their daily lives are compared. The context sources leading to the best keyframe selection are summarised. Figure 5.1 shows the position of this chapter’s work in the whole model. Figure 5.2 presents the process to segment events.
Figure 5.1: Work in Chapter 5
Source: The author (2013) Process
Data collection
(Multiple sources; multiple time points) Data analysis Detecting contexts Segmenting events Generating narratives Data output/display
(Different output formats based on different devices)
Data Storage
114
Figure 5.2: The process to segment events
Source: The author (2013) As shown in Figure 5.2, the event segmentation includes two key parts namely attribute extraction and boundary detection. Extracting attributes needs detecting the change of context, and standardising unit due to the different frequency in data sources. We begin with the change of context as follows.