AutoAct: An Auto Labeling Approach Based on Activities of Daily Living in the Wild Domain
Iqbal Hassan, Abtahi Mursalin, Robin Bin Salam, Nazmus Sakib, H M Zabir Haque · 2021
Annotation problem in Human Activity Recognition (HAR) has always been a major concern. The main problem of human-based annotation is that it often leads to misclassification due to uneagerness, hectic working and living conditions, misleading emotional fluctuation, etc. A solution for that can be the use of a fully automated annotation system. Here, we proposed a unique method for developing a pre-trained model which will be used as the auto annotation tool, using two completely different benchmark Activities of Daily Living (ADL) in the wild data sets. Our proposed auto labeling system has been trained in varied conditions which reduce the bias from location, people, and device, and makes our system very robust. Here, we used a unique mixture of augmentation, scaling, and boosting to increase accuracy, reduce overfitting, and handle imbalanced class problems. The accuracy achieved for auto annotation in case of data set 1, data set 2, and combined data set are 92%, 87%, and 73% respectively. Also, performance evaluation has been done using precision, recall, f1 score, and ROC curve, and the results achieved are promising.