Doctoral Thesis
Refine
Document Type
- Doctoral Thesis (3) (remove)
Language
- English (3) (remove)
Has Fulltext
- yes (3)
Keywords
- Machine learning (3) (remove)
Faculty / Organisational entity
The neural networks have been extensively used for tasks based on image sensors. These models have, in the past decade, consistently performed better than other machine learning methods on tasks of computer vision. It is understood that methods for transfer learning from neural networks trained on large datasets can reduce the total data requirement while training new neural network models. These methods tend not to perform well when the data recording sensor or the recording environment is unique from the existing large datasets. The machine learning literature provides various methods for prior-information inclusion in a learning model. Such methods employ methods like designing biases into the data representation vectors, enforcing priors or physical constraints on the models. Including such information into neural networks for the image frames and image-sequence classification is hard because of the very high dimensional neural network mapping function and little information about the relation between the neural network parameters. In this thesis, we introduce methods for evaluating the statistically learned data representation and combining these information descriptors. We have introduced methods for including information into neural networks. In a series of experiments, we have demonstrated methods for adding the existing model or task information to neural networks. This is done by 1) Adding architectural constraints based on the physical shape information of the input data, 2) including weight priors on neural networks by training them to mimic statistical and physical properties of the data (hand shapes), and 3) by including the knowledge about the classes involved in the classification tasks to modify the neural network outputs. These methods are demonstrated, and their positive influence on the hand shape and hand gesture classification tasks are reported. This thesis also proposes methods for combination of statistical and physical models with parametrized learning models and show improved performances with constant data size. Eventually, these proposals are tied together to develop an in-car hand-shape and hand-gesture classifier based on a Time of Flight sensor.
Regular physical activity is essential to maintain or even improve an individual’s health. There exist various guidelines on how much individuals should do. Therefore, it is important to monitor performed physical activities during people’s daily routine in order to tell how far they meet professional recommendations. This thesis follows the goal to develop a mobile, personalized physical activity monitoring system applicable for everyday life scenarios. From the mentioned recommendations, this thesis concentrates on monitoring aerobic physical activity. Two main objectives are defined in this context. On the one hand, the goal is to estimate the intensity of performed activities: To distinguish activities of light, moderate or vigorous effort. On the other hand, to give a more detailed description of an individual’s daily routine, the goal is to recognize basic aerobic activities (such as walk, run or cycle) and basic postures (lie, sit and stand).
With recent progress in wearable sensing and computing the technological tools largely exist nowadays to create the envisioned physical activity monitoring system. Therefore, the focus of this thesis is on the development of new approaches for physical activity recognition and intensity estimation, which extend the applicability of such systems. In order to make physical activity monitoring feasible in everyday life scenarios, the thesis deals with questions such as 1) how to handle a wide range of e.g.
everyday, household or sport activities and 2) how to handle various potential users. Moreover, this thesis deals with the realistic scenario where either the currently performed activity or the current user is unknown during the development and training
phase of activity monitoring applications. To answer these questions, this thesis proposes and developes novel algorithms, models and evaluation techniques, and performs thorough experiments to prove their validity.
The contributions of this thesis are both of theoretical and of practical value. Addressing the challenge of creating robust activity monitoring systems for everyday life the concept of other activities is introduced, various models are proposed and validated. Another key challenge is that complex activity recognition tasks exceed the potential of existing classification algorithms. Therefore, this thesis introduces a confidence-based extension of the well known AdaBoost.M1 algorithm, called ConfAdaBoost.M1. Thorough experiments show its significant performance improvement compared to commonly used boosting methods. A further major theoretical contribution is the introduction and validation of a new general concept for the personalization of physical activity recognition applications, and the development of a novel algorithm (called Dependent Experts) based on this concept. A major contribution of practical value is the introduction of a new evaluation technique (called leave-one-activity-out) to simulate when performing previously unknown activities in a physical activity monitoring system. Furthermore, the creation and benchmarking of publicly available physical activity monitoring datasets within this thesis are directly benefiting the research community. Finally, the thesis deals with issues related to the implementation of the proposed methods, in order to realize the envisioned mobile system and integrate it into a full healthcare application for aerobic activity monitoring and support in daily life.
In one-dimensional (1-D) Ultrasound (US) measurements, signals are
acquired that form the basis of more sophisticated two-dimensional (2-D) or
three-dimensional (3-D) US imaging. These 1-D signals contain a lot of raw
information about the US wave propagation and interaction with the
medium that is only processed in parts during image generation. While
image representations are easy to interpret for humans, the analysis of US
wave signals is hard to perform without applying algorithms to extract
desired features.
This work investigates reliable and fast 1-D US signal classifications to
distinguish between different stages or states in biomedical US scenarios and
shows how the new field of Machine Learning (ML) on raw US wave data
provides advantages and different applications. To achieve good results, the
input signals are treated as time series, which requires the deployment of
comparatively complex Time Series Classification (TSC) algorithms.
The literature shows that a lot of research efforts have previously only
tackled the classification and segmentation of US Brightness mode (B-Mode)
images, while neglecting approaches to classify 1-D signals to a large extent.
This research contributes by developing, deploying and evaluating
classification approaches for three distinct biomedical US classification tasks
and finds that respective signal classifications for different scenarios are
possible with varying degrees of accuracies. It entails the comparison of
several combinations of data types (e.g. temporal, spectral and statistical
features or raw signals), ML models and pre-processing steps to provide a
strong foundation for robust, binary classifications of 1-D US signals for
scenarios based on low-cost wearable, mobile and stationary devices. This
research addresses scientific questions not answered before by informing on
detailed descriptions of beneficial domain specific knowledge (domain specific
knowledge (DSK)), achieved accuracies and times needed for training and
evaluation of the examined ML models.
The resulting ML pipelines includes solutions based on data acquired from
custom experimental setups or clinical trials. Possible real-world applications
might include muscle contraction trackers, muscle fatigue detectors,
epiphyseal radius bone closure detectors or devices providing information
about advanced liver disease stages.
Automated machine-assisted
classifications requiring as little DSK as possible from the end user enable
application scenarios ranging from fitness or rehabilitation trackers as
consumer devices to solutions providing diagnostic support without requiring
extensive knowledge from professional medical practitioners. For example,
decision support systems for bone age assessments in clinical use or liver
health assessment systems for gastroenterologists.
This work shows that reliable, robust and fast classifications based on 1-D
US signals are possible with high degrees of accuracies depending on the
examined scenario with achieved F 1 -scores ranging from ≈ 70% to ≈ 87%.
These results prove that real-life applications for recreational purposes are
already possible and that critical applications for clinical use are highly likely
to be achieved once the presented approaches are further optimized in the future.