 Welcome and thanks for joining our talk today. We are Anna-Talia Ho-Gestrat and Alexandra Illiger, and we are student research assistants at the Karl von Nussjärzke University of Oldenburg. Together with Jörg-Henrik Bach and Sarah Blum from Hörtech and the University, we investigated an automatic detection of human activities from axillary metasensors integrated in heribals. The motivation for this research is improving a fearing aid performance. With activity tracking, it's possible to have different audio processing for different activities, so the audio processing fits the needs of the user. Recently, meaning the past few years, the field of activity detection through body-worn sensors has been an area of interest for many people. But still, there's not much information about human-worn ear-level devices. For this project, we got research devices kindly provided by Doppel, which have in real-time access to a three-channel axillary metasensor data. As shown on the slide, the gathering of the data is done by performing an activity, while the data is sent via Bluetooth from the heribal to the computer, and then an activity label is added and a CSV file is produced. Due to corona, our database is not as extensive as we would like it to be, but for now we have two sets of data, meaning we have two data sets from two individuals performing 12 different activities for five minutes each with a sampling rate of 800 Hz. On the left you can see the list of our investigated activities, and on the right you can see some example data. Our activities include whole body movements such as walking or jumping, but also more subtle activities like reading or typing. On the right, we brought the graphs of three activities of one of the three axillary metas channels and with the data for about two minutes. You can already see that the three activities look pretty different on the graphs. Our data processing takes place in three steps. Pre-processing, feature extraction and classification. For pre-processing, we are removing the outliers defined as three times the standard deviation, meaning that the whole sample is removed when there is an outlier detected for one of the three channels. Then we balance between activities, so we have same amount of data for each activity again, and then the data is scaled between zero and one for each activity so that the highest value becomes one and the lowest is zero. For feature extraction, we calculate the mean and the standard deviation over a window of one second. For classification, we use the Gaussian knife-based classifier. The Gaussian knife-based classifier can learn a distribution for every feature we extract from our data. In this example here, we have the activity running and so far we extract the mean value and the standard deviation from our time series data. During the training procedure, we then learn our distribution for the features. For a new data point, we decide based on the likelihood to which class they now belong. To do so, the same features are extracted, so again the mean and the standard deviation, and they are compared to the learn distribution from the step four. As already mentioned, this is done for every feature independently, which is a strong and naive independence assumption on our features. To validate our model, we use five-fold cross-validation to make sure that every activity appears at least once in training and in test data. This is what you can see here in this figure. For every class, we have a certain amount of training data and testing data. In particular, we use the so-called stratified keyfold. This is a method for time series data. As for every iteration, a block of corresponding features is extracted. In our case, this block may consist of five consecutive means, for example. As a result for our two participants, we achieved an overall accuracy of 67% for the first one and 72% for the second participant. Here on this slide, you can see the confusion matrices of both participants with the two labels on the y-axis and the predicted ones from our classifier on the x-axis, and every second label has been left out. In both cases, it can be nicely seen that here on the diagonal that most predictions are correct and only for talking, which is displayed here. The predicted label deviates strongly. As for the first one, here we have no distinct classification. While for the second one, it is most of the time identified here as reading. So far, we achieved good results and there are still a lot of possibilities to improve and to gain even better results in the future. Even with that small amount of data we had due to COVID, it can be stated that the accelerometer data works for activity detection with simple machine learning models, feature extraction and pre-processing. To optimize the results even more, we will collect more data to optimize and generalize the pre-processing feature extraction and the classification. But so far, the overall pipeline is ready. Moreover, we have the advantage that we are using simple machine learning, which provides the opportunity to change the activity detection from an offline to an online approach on, for example, mobile devices. And in general, the area of possible applications is huge. So not only in hearables, the detection would be a smart gadget to, for example, automatically reduce the music level when starting talking to someone. For now, we investigated the activity recognition using these hearables. But it can also be easily transferred to hearing aids as well. And in hearing devices, this could be a really useful and powerful tool to maybe activate different signal processing algorithms according to specific activities. This may be when you are jogging, you want to have a broad field of acoustic attention as you want to hear a car from behind as well as from the side. When you're cycling, you don't want to have the wind feedback in your ears. And when you start talking to someone, you may maybe want to activate a directed beamformer or stuff like that. Yeah, so therefore, this is a really interesting field of research with a really promising future. And we're glad that we have the possibility to work on such a project. In that context, we would like to thank our founders to give us the possibilities for our research. And we want to thank Doppel for providing us the research devices.