View Item 
        •   Utrecht University Student Theses Repository Home
        • UU Theses Repository
        • Theses
        • View Item
        •   Utrecht University Student Theses Repository Home
        • UU Theses Repository
        • Theses
        • View Item
        JavaScript is disabled for your browser. Some features of this site may not work without it.

        Browse

        All of UU Student Theses RepositoryBy Issue DateAuthorsTitlesSubjectsThis CollectionBy Issue DateAuthorsTitlesSubjects

        Video-based Activity Recognition for Child Behaviour Understanding

        Thumbnail
        View/Open
        Masters_Thesis_Feyisayo_Olalere_Submission.pdf (1.939Mb)
        Publication date
        2021
        Author
        Olalere, F.E.
        Metadata
        Show full item record
        Summary
        Over the years, deep learning models have been able to record state-of-the-art (SOTA) performance on the task of activity recognition. The results of this can be seen in applications such as video surveillance, medical diagnosis, robotics for human behavior characterization, and like in this study, recognition of human activities from videos. One of the factors that have contributed to the benchmark performance of these models is the availability of large-scale datasets. However, we have observed that these datasets are largely skewed towards adults. That is, they contain more videos of adults than kids. Out of 5014 videos from an adult-specific dataset, only 1109 videos contained kids performing an action. Since there are visual differences in how an adult performs an activity instead of a child, in this study, we test if current SOTA deep learning models have some systemic biases in decoding the activity being performed by an adult or a kid. To do this, we create kid-specific and adult-specific datasets. Using a SOTA deep learning model trained on the different datasets, we test for the generalization ability of the deep learning model. Our results indicate that, while SOTA deep learning models can be used to classify kid activities, the kid-specific dataset is more complex to generalize to than the adult-specific dataset. The study also shows that the features learned from training on a kid-specific dataset alone can be used to classify adult activities while the reverse is not the case.
        URI
        https://studenttheses.uu.nl/handle/20.500.12932/39771
        Collections
        • Theses
        Utrecht university logo