You are in: results>D3.3
Results
-----------------------------------------------------------------------------------------------------------
Deliverables
-----------------------------------------------------------------------------------------------------------
D3.3-Report on outdoor recontruction and
interpretation techniques
Considering that outdoor Confidence localization system will provide not only location data, but also data from accelerometers, gyroscopes and magnetic field sensors, more related work on fall detection and activity recognition becomes relevant than it used to be for the indoor version. Fall detection with accelerometers and gyroscopes can reach very high accuracy even with simple threshold-based algorithms. Therefore we believe these sensors will be valuable for Confidence and we will try to exploit them through machine learning and threshold rules. It should be noted, though, that experiments with these sensors reported in the literature were performed on types of falls and activities of daily living likely to be distinguished by acceleration and velocity, whereas in Confidence we hope to tackle more difficult falls as well.
The first step to be performed by the reconstruction and interpretation system is preprocessing, whose goal is to remove the noise in the data from the localization system. It removes constant small errors with Kalman filter and short-term large errors with median filter. More problematic are long-term large errors, which we plan to remove by a filter that enforces anatomic constraints, i.e., it recognizes when the user’s perceived posture is anatomically unlikely. We also intend to improve localization by fusion with other sensors.
The next step is activity recognition, which will mostly be carried out by a machine-learned classifier. Its attributes will be distances between body tags, tag velocities, angles between body parts, accelerations, angular velocities and orientations of tags with respect to the Earth’s magnetic field. Short intervals of time will be classified into activities and several ways to merge attributes derived from consecutive sensor readings in those intervals into attribute vectors are considered. We also plan to use expert rules and possibly other methods. The advantage of expert rules is that an expert understands what the general properties of an activity are and can ignore properties specific to an execution of the activity. Furthermore, an expert can imagine variations of the activity. We will pay special attention to user adaptation. We hope to use one of semi-supervised machine learning methods, which improve classification by combining labelled and unlabelled data.
After activity recognition, fall detection is performed. This step uses the information about falling, lying and other postures in which the user may end up after a fall that were recognized during the previous step in order to decide whether a fall has occurred. For this purpose we plan to use expert rules whose parameters will be tuned automatically.
The final task of the reconstruction and interpretation system is general disability/disease detection. It consists of two parts: analyzing the movement of the user’s body parts with respect to each other (micro movement) and the user’s movement from one location in their environment to another (macro movement). The user’s micro movement will be analyzed through statistics characterizing their gait and speed of movement. An outlier detection method will be applied to them, which will learn normal behaviour and detect behaviours sufficiently differing from it. The analysis of the user’s macro movement will first require identifying the locations the user frequents. Then it will track the frequencies of visiting these locations and of various activities performed there and raise a warning should they be unusual.
.
