Activity Identification and Recognition in Real-Time Video Data Using Deep Learning Techniques

Human Activity Recognition is the task of identifying and recognizing the activities performed by a user based on the input received through a sensor. HAR can either be performed by using motion-based sensors for tracking the movement of the person or by capturing the activities performed through a video. This paper focuses on Video Classification-based Human Activity Recognition. The Human Activity Recognizer takes the video consisting of a sequence of frames as input and predicts the activity being performed in the video. The state-of-the-art video classifiers for Human Activity Recognition like Baseline Single Frame architecture, Late Fusion, Early Fusion, Slow Fusion, MoViNet (3D CNN), Long Range Convolutional Network and Slow Fast Network are discussed and implemented. The architectures are built using TensorFlow. The architectures are trained and evaluated on the UCF–101 data set. The single frame architecture is a strong baseline for HAR and gives an accuracy of 94.26%. The MoViNet and Slow Fast Network gives an accuracy of 95.24% and 96.21%, respectively.