Video Categorization using Feature and Context Extraction

Video classification has received a great deal of attention in machine learning because of its wide use in a number of crucial applications, including human activity recognition and dynamic scene layout. However, video arrangements show great vacillation as a result of large size changes and camera movement, posing serious difficulties for both video representations and classification. Video information must be processed significantly more thoroughly than other types of data, which is its main disadvantage when used for classification, also in terms of storage it is to be noted that these videos do not come in small sizes. Deep learning is one of the most fitting solutions to this problem. Models like Long Short-Term Memory (LSTM) and Convolutional Neural Networks (CNN) are one of the most significant approaches to video level categorization alongside Spectrogram for audio level categorization. This paper proposes an ensemble solution to the problem put forth by the Department of Space, Indian Space Research Organization (ISRO) in Smart India Hackathon 2022 to categorize videos and short clips of ISRO dataset. A multi modal approach has been put forth in this paper, which involves feature extraction and context extraction using a variety of datasets (Audio dataset, ISRO video dataset and YouTube 8M dataset).