Comparative study of CNN, VGG16 with LSTM and VGG16 with Bidirectional LSTM using kitchen activity dataset

Computer-based human activity detection of everyday life has lately attracted a lot of attention due to its relevance to contextual assisted living. Such applications necessitate the automatic recognition of high-level events involving many human actions in a specific context. Kitchen activity recognition aims to monitor the activities of the inhabitant and identify deviations from the normal in kitchen. This paper presents a deep neural architecture for recognizing kitchen activities that employs a combination of machine learning models. In this work, comparative study of kitchen activity detection is performed using kitchen video dataset with three models, namely CNN, VGG16 with LSTM and VGG16 with Bidirectional LSTM. The results of the experiments reveal that an accuracy of 31% for training and 28% for testing is achieved with CNN model, whereas VGG16 with LSTM model achieves 92% for training and 96% for testing, VGG16 with Bidirectional LSTM model achieves 95% for training and 94% for testing.