Encapsulated Features with Multi-objective Deep Belief Networks for Action Classification

Human action classification plays a challenging role in the field of robotics and other human-computer interaction systems. It also helps people in crime analysis, security tasks, and human support systems. The main purpose of this work is to design and implement a system to classify human actions in videos using encapsulated features and multi-objective deep belief network. Encapsulated features include space-time interest points, shape, and coverage factor. Initially, frames having actions had been separated from the input videos by means of structural similarity measure. Later, spatiotemporal interest points, shape and coverage factor are extracted and combined to form encapsulated features. To improve the accuracy in classification, MODBN classifier was designed by combining multi-objective dragonfly algorithm and deep belief network. Datasets such as Weizmann and KTH are used in MODBN classifier to carry the experimentation. Accuracy, sensitivity, and specificity are measured to evaluate the classification network. This proposed classifier with encapsulated features can produce better performance with 99% of accuracy, 97% of sensitivity, and 95% of specificity.