Chen Jimena Melisa Parodi 302424544 Menashe Shalom 301376869 Allowing Early Inspection of Activity Data from a Highly Distributed Bodynet with a Hierarchical-Clustering-of-Segments Approach Chen Jimena Melisa Parodi 302424544 Menashe Shalom 301376869

Body-wide inertial sensing systems Purpose: The output delivered by body-wide inertial sensing systems has proven to contain sufficient information to distinguish between a large number of complex physical activities. מערכת סנסורים הפרושים על הגוף לאפיון תנועה.

Body-wide inertial sensing systems Problem: The high dimensionality of the raw sensor signals with the large set of possible features tends to increase rapidly causing a problem in the parts that calculate and select features.

Body-wide inertial sensing systems Existing methods for reducing large input dimensions- traditional feature selection methods such as: Boosting-based approach Component Analysis Wrapper methods

Body-wide inertial sensing systems New proposed method: Hierarchical-Clustering-of-Segments Approach

Hierarchical-Clustering-of-Segments Approach Relies on the hierarchical clustering of discovered patterns of both inertial trajectories and angular data, across all of the body-worn sensors. Allows thorough human inspection of which data segments are discovered as meaningful for given activities.

Hierarchical-Clustering-of-Segments Approach Five phases of the proposed system

Hierarchical-Clustering-of-Segments Approach The contributions: Showing that it is possible to obtain an early insight into which sensors and features can be expected to perform well in distinguishing the different target classes.

Hierarchical-Clustering-of-Segments Approach Showing that it is possible to inspect at an early stage which classes are prone to misclassification between each other for particular sensors and features.

Hierarchical-Clustering-of-Segments Approach Showing that many problems appearing at classification can be identified already at the clustering stage.

Hierarchical-Clustering-of-Segments Approach Experiment and methodology Dataset 19 activities from a car maintenance scenario, workers wearing a body sensor network integrated in a jacket with sensors in their torso, upper an lower arms and hands.

Hierarchical-Clustering-of-Segments Approach Experiment and methodology

Hierarchical-Clustering-of-Segments Approach Experiment and methodology Visual Inspection at the clustering level 1. How can the high dimensional data be analyzed at the clustering level for its suitability for activity recognition, without resorting to feature selection?

Hierarchical-Clustering-of-Segments Approach Experiment and methodology Visual Inspection at the clustering level 2. How do the generated clusters correlate to the final activity recognition results?

Hierarchical-Clustering-of-Segments Approach Experiment and methodology Visual Inspection at the clustering level Bright colors- High similarity. Diagonal white line- Cluster’s precision value of pairs of the same class. Subplots- symmetric divided by the diagonal white line.

Hierarchical-Clustering-of-Segments Approach Experiment and methodology Visual Inspection at the clustering level The visual Inspection showed that it is possible to: Identify channels which tend to mix up different clusters from those who have a good class distinctiveness. See for single channels which classes data tend to be close to each other and might get easily mixed up.

Hierarchical-Clustering-of-Segments Approach Experiment and methodology Comparison with the Recognition results The ability to detect classes that are hard to distinguish at the clustering phase would not only lead to a faster way of inspecting the data: we could also examine which sensors (or body locations) would contribute to such difficulties.

Algorithm description Problem dimensionality of the resulting data Solution concentrate only on the important signals concentrate only on parts of the data that are significant for the activity

Algorithm description Maybe dig more about the purpose of the alg.

Algorithm description Data preprocessing and segmentation Clustering Cluster Analysis Classification

Algorithm description Data preprocessing and segmentation data from the sensors is used to construct a body model Low pass filtering to segment the continuous signal להסביר קצת על חיישני כיוון וחיישני זיויות וכו'.

Algorithm description B. Clustering Total of 34 channels Goal – find significant channels and segments for each activity we want to spot Hierarchical clustering Dynamic Time Warping (DTW)

Algorithm description C. Cluster Analysis To get an idea of the distinctiveness between classes of single channels we calculate the cluster precision on the cluster.

Algorithm description C. Cluster Analysis Precision – Take all the clusters of one channel and one class and find clusters centers Calculate distances between cluster centers in both classes using dynamic time warping. The sum of the smallest distances divided by the number of clusters provides the cluster precision.

Algorithm description C. Cluster Analysis The higher the Precision the more dissimilar are the cluster center of the two classes on this channel Figure 2 – white = similar Black = different

Algorithm description C. Cluster Analysis Examples – Z axis of the hand tip, wrist and also of the elbow have a good distinctiveness shoulder is not very significant

Algorithm description Classification Calculate distances between test data segment and all cluster centers. Found activity Thresholds – DTW where 95%of the class members still nearer to cluster center Mean of all DTW from cluster members to cluster centers.

Algorithm description Classification Channels are sorted by summing all cluster precision values for each channel (Table 2)

OBSERVATIONS AND DISCUSSION Cluster precision insight of channel performance determine which channel has the most distinguishable character between the different classes up and down movements seem to be quite characteristic for the activities of our dataset worst channels are the shoulder in y and z direction Only the x axis of the shoulder ranks higher, demonstrating that leaning forward and backward can give hints to some of our activities.

OBSERVATIONS AND DISCUSSION B. Cluster precision insight of class performance Classes can Easley mix up Summing up all cluster precision values of one class on all channels shows close values