Motion Embedded Images: An Approach to Capture Spatial and Temporal Features for Action Recognition

The demand for human activity recognition (HAR) from videos has witnessed a significant surge in various real-life applications, including video surveillance, healthcare, elderly care, among others. The explotion of short-form videos on social media platforms has further intensified the interest in this domain. This research endeavors to focus on the problem of HAR in general short videos. In contrast to still images, video clips offer both spatial and temporal information, rendering it challenging to extract complementary information on appearance from still frames and motion between frames. This research makes a two-fold contribution. Firstly, we investigate the use of motion-embedded images in a variant of two-stream Convolutional Neural Network architecture, in which one stream captures motion using combined batches of frames, while another stream employs a normal image classification ConvNet to classify static appearance. Secondly, we create a novel dataset of Southeast Asian Sports short videos that encompasses both videos with and without effects, which is a modern factor that is lacking in all currently available datasets used for benchmarking models. The proposed model is trained and evaluated on two benchmarks: UCF-101 and SEAGS-V1. The results reveal that the proposed model yields competitive performance compared to prior attempts to address the same problem.