Hidden Markov Model-Based Smart Annotation for Benchmark Cyclic Activity Recognition Database Using Wearables

Symbolic picture for the article. The link opens the image in a large view. — Layout of the activities (Germany). The starting point of each activity is shown by the letter whose key is in the table on the right. Double lines in the table separate the 4 sections of activities.

24. April 2019

Activity monitoring using wearables is becoming ubiquitous, although accurate cycle level analysis, such as step-counting and gait analysis, are limited by a lack of realistic and labeled datasets. We address this need for both a benchmark dataset as well as the smart annotation methods used to create it in our new paper Hidden Markov Model-Based Smart Annotation for Benchmark Cyclic Activity Recognition Database Using Wearables authored by Christine Martindale, Sebastijan Sprager and Bjoern Eskofier.

The open access article published in Sensors can be found here: https://www.mdpi.com/1424-8220/19/8/1820

Dataset is publicly available here: https://www.mad.tf.fau.de/research/activitynet/benchmark-cyclic-activity-recognition-database-using-wearables/

Sensor type and location. (a) Photograph of IMU sensor system in 3D printed case. (b) Diagram showing sensor attachment to shoe, using industrial Velcro and Moticon insole, which was used instead of the original sports shoe insole pressure sensor location and approximate size and shape within the insole. Also showing axes location for the insole accelerometer. (c) IMU sensor locations on the body with corresponding axes.

Layout of the activities (Germany). The starting point of each activity is shown by the letter whose key is in the table on the right. Double lines in the table separate the 4 sections of activities.

Logarithmic histogram of walking bout lengths, for all subjects.

Boxplot of the stride times per activity, over all subjects.

Boxplot of the swing duration per activity, over all subjects.

Boxplots of the effort required for manual correction of each batch of subjects using the smart annotation approaches. (a) Shows the actual effort per batch where Batch 1 was using the edge-detection method and Batches 2 to 7 were using the hHMM method, where the data from the previous batches is always added to the training set of the subsequent model. (b) Shows the labeling effort for each batch’s hHMM model when predicting the subject data from Batch 7.