Handling Imbalanced Data Using a Cascade Model for Image-Based Human Action Recognition

Human action recognition plays a crucial role in intelligent monitoring systems, which are based on analyzing the possibility of anomalous events related to human behavior, such as theft, fights, and other incidents. However, by definition, anomalous events occur somewhat infrequently, thus leading to small and unbalanced data compared to data on other events. Such a data imbalance causes the human action recognition model to fail to produce optimal accuracy. To overcome the problem of imbalanced data, the typical methods used are oversampling and undersampling. However, these two methods are not considered to be very effective, because they cause the loss of a significant amount of information or deviations from reality. Therefore, the current paper proposes a cascade modeling strategy to address data imbalance problems, particularly in the case of human action recognition. The proposed strategy consists of several steps: preprocessing, feature extraction, modeling, and evaluation. The BAR dataset experiment found that the cascade model outperformed the other examined methods with an accuracy of 56.38%. However, there is still potential for further improvement through continued research.