Comparison of LSTM and GRU in Video Activity Recognition Using Transfer Learning and Time Distributed Layer in Keras

With the advancements in Deep Learning, many feats are achievable today that can be considered technologically appealing and also very useful in real life. Our paper aims to recognize human activity in videos using Deep Learning and compare Gated Recurrent Unit (GRU) efficiency with Long Short Term Memory (LSTM) and various transfer learning models to determine an optimal and simplistic action recognition mode. Taking video as input, we convert it into sequences of frames, process each frame through a Convolutional Neural Network (CNN), and connect the entire sequence, using a time distributed layer, to LSTM or a GRU. To achieve this, we train our model on UCF101 and HMDB51, two large video activity datasets. UCF101 has 101 action classes that are widespread across various activities, and HMDB51 has 51 action classes that are more focused on humans' movement. Activity recognition is beneficial as it can help us refine many parts of society, especially post-pandemic. This project is our attempt at activity recognition and analysis on a couple of methods that can hopefully be helpful.