Capítulo 3 El aprendizaje cooperativo en la enseñanza del inglés como lengua extranjera.
3.3 Factores que definen una situación comunicativa
This chapter has demonstrated that behaviour recognition has a wide range of appli- cations and has been drawing significant attention from the machine research commu- nity. Generally, all past works demonstrated that behaviours can be inferred from a sequence of sensor observations output from the house and that a supervised learning method is often used to recognise the behaviours, where the algorithm is trained from a set of labelled sensor data.
This chapter also reviewed the range of sensing methods that could be used to recognise the inhabitant’s behaviours from the home, ranging from cameras and mi- crophones to wearables, motion and objects-used sensors. The choice of which sensing method to use depends upon what the aim of the smart home is, the level of infor- mation obtained from the sensors and the level of privacy intrusion.
There have been a lot of smart home research projects reported in the literature and some of them have been reviewed. Since each of these projects has its own directions and aims, comparisons between the different smart home projects are very difficult. However, the literature on these smart home projects has demonstrated the need for further machine learning methods to learn the behaviours of the inhabitant, and there is still far more work that needs to be done in this area before the smart home is able to accurately monitor the inhabitant’s behaviours.
As this will become clearer in later chapters, using state-change sensors for be- haviour recognition raises one problem, i.e., the series of sensor readings output from
the home is unsegmented, where the start and end boundary of a behaviour is un- known. Very few methods described in the literature perform segmentation and behaviour recognition simultaneously on the sensor stream. Addressing this is impor- tant in real implementations as the sensor stream needs to be segmented before any classification is performed.
The literature has demonstrated that supervised learning has been the primary approach for behaviour recognition. However, it needs labelled data. Although there are unsupervised methods proposed in the literature, they do not satisfy the criteria that are considered in this thesis, which are described Section 1.4.2.
In the next chapter, we describe one particular behaviour recognition problem that still deserves further attention and our proposed method to approach the problem.
Chapter 3
A Supervised Learning Approach
to Behaviour Recognition
This chapter describes a supervised learning approach to behaviour recognition and the segmentation of the sensor stream. The behaviour recognition problem is first described, and then an overview of methods that have been proposed in the literature is provided. An approach to the problem based on simultaneous segmentation and recognition is presented, together with experiments showing the effectiveness of the proposed method for recognising human behaviour.
3.1
Problem Description
As described in Section 1.4.1, the smart home uses sensors to collect information about the inhabitant’s activities. Given that we have a sequence of tokens obtained from the sensors, the question is how to recognise behaviours. The challenges in this task are that behaviours are rarely identical on each use; the order in which the individual components happen can change; and components can be present or absent at different times (for example, making a cup of tea may involve milk, or may not, and the milk could be added before or after the hot water). Adding in the fact that the activities are not directly observed, and that sensor observations are themselves intrinsically noisy, the tokens represent only a very partial picture of what the inhabitant is doing.
One common approach to recognising behaviours is to use a supervised learning method. In supervised learning, the algorithm is presented with a set of labelled behaviours that are to be learned. The algorithm is then trained to minimise some error norm between the predicted output of the classifier and the target labels. In this way it should be able to generalise to other, previously unseen, inputs. The benefits of this are a clear delimitation of the problem, and access to many algorithms in machine learning, including decision trees, statistical classifiers such as the na¨ıve Bayes classifier, and graphical models such as the hidden Markov model (HMM). And indeed, there are several examples of these methods implemented in the various smart homes that have been reported in the literature, such as [136, 133, 151].
However, as described in Section 1.3, the sensory stream in a smart home consists of an unending series of sensor readings, which poses an additional challenge to seg- ment the sensory stream in order to determine the start and end points of an activity boundary, before any classification can be performed on the sensor sequences that represent individual behaviours.
Most current approaches assume that the activities have been segmented, or use a fixed window length to partition the input stream. As each behaviour can be described by different numbers of sensor readings, it is inappropriate to rely on a fixed window length, since it is unlikely that all the sequences in the window belong to one behaviour. We want to ensure that other behaviours in the window are also recognised. We approach this problem by proposing a method using a set of hidden Markov models that each recognises a behaviour and compete to explain the current observations, and a method that has the ability to self-determine the window size based on the sensory data.
We first provide an overview of the supervised learning algorithms and describe how they are used to recognise behaviours. We then review methods that are used for activity segmentation. Following this, we introduce our method, which can reliably
detect and segment the sensor stream into behaviours. We use a real smart home dataset to demonstrate the efficacy of our algorithm.