3. MATERIALES Y MÉTODOS
3.1. ELECTRODIÁLISIS
3.1.2. Proceso estándar de electrodiálisis
dataset, the combination of HMM and SVM outperformed the average classifi-cation recalls of using supervised machine learning algorithms alone. Another experiment conducted on vehicle in Chapter 7 also confirmed that the hidden Markov model can effectively improve the reliability of environment detection.
9.1.4 Context association
If the behaviours and environments are not independent in reality, how can they be associated? How can context association be used to reduce the chances of the context determination algorithms selecting an in-correct context?
The purpose of context association is to improve the reliability of context determination. It may be used in two ways, by either improving behaviour de-tection from environment information, or the reverse. As all behaviour categories covered in this study can appear in every environment, improving the reliability of environment detection with the aid of behaviour information was focused in Chapter 7.
It was found that the behaviour information could be exploited in two dif-ferent ways for enhancing context determination. First, the environments can be better predicted according to whether the behaviour is static or not. Sec-ond, different features and classification models can be implemented depending on whether the user is on a pedestrian or on a vehicle. The analysis of the ex-periments under different scenarios have confirmed that each of the association methods can effectively reduce the number of incorrect environment detection and improve the reliability of environmental context determination.
9.1.5 Demonstration of context-adaptive navigation
How can context determination be implemented for context adaptive navigation? What improvements can context adaptations bring for a navigation system?
The contribution of context adaptations has been demonstrated in Chap-ter 9 by a tentative experiment conducted on pedestrian across both indoor and outdoor environments. The navigation system selected different positioning tech-niques according to the detected environmental contexts. In the detected indoor environments, the PDR algorithm was applied; while in the outdoor environment, GNSS positioning results were implemented. By comparing the positioning per-formances using PDR alone, GNSS alone and context adaptive navigation, it is
168 CHAPTER 9.
shown that using the detected environments to select appropriate navigation tech-niques can improve both the availability and accuracy of positioning solutions.
9.2 Recommendations
Based on the research investigated in this thesis, future work could carry on for better context determination and positioning performance. They have been addressed in Section 9.2.1 and 9.2.2 respectively.
9.2.1 Further research
The following research arising from the limitations of the work in this thesis could be extended or improved in the future.
1. Implement deep learning for behaviour detection. Chapter 4 demonstrated the behaviour recognition by using some typical supervised machine learning algorithms. The structure of these algorithms are simple and explicit, and the input features have to be manually extracted. However, as the amount of data increases, the performance of these learning algorithms, like SVM and decision trees, does not improve a lot. They tend to plateau after a certain training point (LeCun et al.,2015). On the contrary, the deep learn-ing method is able to learn from dense and complex hierarchical networks that transform the raw data (e.g. image, voice, text and sensor signal) into inferences/predictions. The structure of a generic deep learning architecture is presented in Figure9.1. Moreover, it can learn feature representations di-rectly from raw data rather than relying on domain-specific features. Deep learning approaches have shown better generalization ability than shallow methods and widely applied on many classification tasks, such as computer vision, speech recognition, natural language processing and bioinformatics (Ciregan et al., 2012; Krizhevsky et al., 2012). With the high-performance CPU and GPU deployed within a smartphone, the heavy computation load of deep learning could be handled. This would promote the deep learning approach to achieve better classification performance for behaviour recogni-tion on smartphone.
2. Use multiple sensors for more robust environment detection. Chapter 5 and Chapter6 have focused on environment detection using GNSS signals.
However, the use sometimes prefers to switch off the GNSS module, and the GNSS based approach may provide misleading results before a cold start has completed. Further improvement can be considered by integrating other
sen-9.2. RECOMMENDATIONS 169
Inference Sensor
Data
Input Layer Hidden Layer Output Layer
Figure 9.1: A typical deep neural network structure
sors for environment detection. As proposed in Li et al. (2014) and Groves et al.(2013b), light sensor, magnetometers, cellular and Wi-Fi signals could all be potentially useful for environment detection. Their pros and cons have been discussed in Chapter 3. Although individually, they cannot pro-vide a reliable environment prediction, their integration along with GNSS signals may further improve the reliability of environment detection and of-fer a backup system when the GNSS module of the smartphone is switched off by the user. In terms of the detection framework, the factorial hidden Markov model or LSTM (Long Short-Term Memory) network are the po-tential options that can be considered to extend the current framework and to combine sensor measurements or individual prediction for determining environments.
3. Enhance context determination by location-dependent connectivity. For temporal connectivity, a time-domain filter has been developed for behaviour recognition in Chapter4where the connection parameters are fixed. Besides temporal relationship, the spatial information can be considered for connec-tivity as well. In reality, the likelihood of one behaviour transiting to another depend on locations (Groves et al.,2013b). For example, the connection with stationary or moving trains is more likely to happen in the train station.
By exploiting this spatial information, the reliability of context determi-nation should be further enhanced. The location-dependent connectivity relationship could be estimated from GIS data and the uploaded behaviour and location information via crowdsourcing. Similar enhancement can be applied for environment detection as well.
170 CHAPTER 9.
9.2.2 Application of context detection on navigation sys-tem
In this thesis, context determination has been investigated to build the basis for context adaptive navigation. The following work could be conducted for ubiqui-tous positioning and better positioning accuracy across different contexts.
1. Use detected contexts to improve the operation of the orientation sensors.
Currently, the orientation of a smartphone is determined from the magne-tometers and accelerometers according to the Android interface (Google, 2018a). However, their accuracy is poor, as demonstrated by the experi-ment conducted in Chapter 8. The poor accuracy not only strongly limits the positioning performance of the PDR algorithm, but also affects the rel-evant applications relying on orientation outputs. For instance, the wrong orientation output might lead the digital map user to wrong places. For a higher accuracy, the horizontal and vertical plane can be determined from the accelerometer measurements only when a static context is detected. The orientation measurements can be fixed in different ways according to differ-ent behaviours. In addition, an extended Kalman filter may be designed for faster convergence after the smartphone orientations change.
2. Implement of turning detection and map matching to improve PDR algo-rithm. Due to the systematic errors of the consumer-grade inertial sensors on a smartphone, the measured heading angles are prone to drift especially after turning. To minimise such effect, one way is to develop an algorithm for reliable orientation measurements as mentioned above. Another way can be considered by implementing environment constraints. A person can-not walk through walls. Turning typically takes place at the intersection points on a map. Once a turning behaviour is detected by the context de-termination algorithm, the turning location can be matched to the nearest intersection points. Therefore the positioning errors of the PDR algorithm can be mitigated.
3. Implement of open-sky/urban classification for intelligent urban positioning.
Different urban positioning techniques have been developed for better po-sitioning performance under different urban scenarios. For example, it was found that GNSS shadow matching performs better in highly dense urban scenarios while the 3DMA ranging approach performs much better in less dense urban scenarios (Adjrad and Groves,2017b). By integrating them via
9.2. RECOMMENDATIONS 171