Tekele Zemamie Action Recognition in Ethiopia Orthodox Tewahedo Church Using CNN-Bi-LSTM

Human action recognition (HAR) is a computer vision technique used to understand the activity of the action performed in the scene. Computer vision technology has become popular and is applied in various areas like surveillance systems, robot vision, satellite imaging, etc. However, using computer vision in human activity recognition is still challenging due to the dynamic movement of human motion, noise, occlusion, complex background, variable dressing style, varying illumination, and others. Various research has been conducted to address these problems, yet they remain challenging. In this study, we applied human action recognition techniques to interpret the meaning of Tekele Zemamie's actions in the Ethiopian Orthodox Tewahedo Church. Tekele Zemamie is church choirs' most common activity during stage ceremonies and other events. To carry out the research experimental work we have prepared 900 video clip datasets. The dataset was recorded in a 2.5-meter distance using a 48MP Camera. From the dataset, 80% was used for training and the remaining 20% was used to test the model's performance. We applied skip frame selection techniques to a single video by selecting 30 frames and we applied spatial and temporal feature extraction techniques. To overcome the stated problems and to develop the recognition model experimental works have been examined using two deep-learning models CNN_LSTM and CNN_Bi-LSTM, and we achieved 95% and 97% recognition accuracy respectively. From the proposed models, CNN_Bi-LSTM works better than the other model.