Advancing Human Action Recognition: A Hybrid Approach Employing Autoencoder and Transfer Learning Techniques

Interest in human action recognition has grown over the past years within the academic community due to practical applications and prospective influence. The research in this area focuses on an analysis and understanding human behavior, mainly body language and gestures. Such behaviors pose an enormous challenge for deep neural networks due to their intrinsic intricality. The complexity comes from the complexity and diversity of the human movements that should be identified and interpreted. Successes in deep learning have changed associated domains, such as image analysis, object recognition, and speech recognition. A massive diffusion of these technologies followed. However, the activity recognition processes that purport to recognize human activities involve more than just detecting body movements. They require understanding the context of a situation and the intention behind those actions. Gesture can thus be defined as any overt body movement used by individuals to communicate information. It is a natural and intuitive way of communicating. This use of gestures as a mode of human-computer interaction portrays a smooth and user-friendly interface, most beneficial to speech and deaf individuals to interact with different categories of people. Our paper proposes a new model of human action recognition that foots the ground between the autoencoder and transfer learning approaches. In this model, an autoencoder as a feature extractor would be used to obtain some basic properties from the input data. These extracted features are finally fed into a pre-trained VGGNet, one of the well-known deep learning architectures, for classification. Our approach considerably enhances accuracy and performance in human action recognition compared to previous models from the results we obtained from our experimentation. Its effectiveness indicates that our proposed model can drive the industry forward, providing practical, real-world applications.