Video Activity Classification : A Comparative Analysis and Deep Learning Based Implementation
Swanand Purkar, Shriwatsal Patil, Varun Kale, Bhakti D. Kadam · 2024
Video Activity classification is a fundamental task in computer vision with applications ranging from video surveillance, human-computer interaction to healthcare and autonomous robotics. Several techniques have been reported in the literature for the same. In recent years, Deep Neural Networks (DNNs) have revolutionized the activity classification, enabling significant advancements in accuracy and robustness. This paper focuses on the deep learning techniques used for human activity detection and classification in videos. Previously statistical approaches were studied and employed for activity classification in which pattern recognition was majorly used. Recently, supervised learning based techniques are much utilized. This papers reviews the human activity recognition and classifying it with the use of video sequences and static images. The crucial steps in classifying the activities such as preprocessing, segmentation, frame extraction, feature vector extraction techniques, and classification models are also discussed. This paper provides a systematic survey of the state-of-the-art techniques, gaps between recent methods and methodologies in activity classification. The key challenges, datasets, and evaluation metrics are discussed throwing some light on the progress and future directions in the domain. Reviewing various surveys, the Long-Term Recurrent Convolutional Networks (LRCN) approach is implemented in this work. The implementation of LRCN approach shows that the model acquires an accuracy of 84.0476% on the action classes from the UCF-101 action datasets.