Authors - Aye Mya Mya Win, Ah Nge Htwe Abstract - In recent years, optical flow-based deep learning methods have pro vided evidence of impressive performance in recognizing human behavioral movements from video sequences, revealing high applicability for fall detection functions. This paper analyzes GMFlow-based architectures by experimenting with three different approaches that merge TCN, Attention, and CNN compo nents. These methods are GMFlow-TCN, GMFlow-TCN-Attention, and GMFlow-CNN-TCN-Attention. The experiments were executed on URFD Da taset, Le2i Dataset, and a combined, URFD-Le2i dataset to analyze and evalu ate their performance. According to the experimental results, the method that combines GMFlow-CNN-TCN-Attention achieved better performance than the other proposed models. This model obtained test accuracies of 100% on the URFD dataset, 92% on the Le2i dataset, and 94% on URFD-Le2i dataset. These results point out that the presented method is capable of effectively cap turing both spatial features and temporal features required for fall detection. This approach provides useful insights for developing effective real-time vi sion-based fall detection applications.
Open Zoom