Video-based Activity Recognition for Child Behaviour Understanding

Abstract

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.