R. Kavitha  and D. Chitra. An Extreme Learning Machine and Action Recognition Algorithm for Generalized Maximum Clique Problem in Video Event Recognition. Dynamic Systems and Applications 30 (2021) No.8, 1228 – 1249

https://doi.org/10.46719/dsa20213082

ABSTRACT.
Video surveillance has garnered quite substantial focus in the past decades and is a prominent subject of research in computer vision. Identifying the events in surveillance videos is still a big challengee, hugely owing to the excessive intra class differencesof events resulting due to variations in the visual appearance, target motion differences, viewpoint change and temporal changes.  In this proposed technical work, an Extreme Machine Learning (EML) and action recognition schemes are developed for semantic concept detection in unnatural videos. In the first step, surveillance videos are considered as the input. The system designed provides ahighly efficient scheme of video representation employing modified salient dense trajectories. The feature encoding is carried out with the help of Vector of Locally Aggregated Descriptors (VLAD) that can hugely render the computational time to be reduced and gain economic hard disk access, and simultaneously, efficiently minimize the feature information loss and boost the recognition accuracy. After this, the Modified Branch and Bound Method (MBBM) are presented to resolve the Problem of Generalized Maximum Clique. After this, these features are utilized for computing deep features by employing Convolutional Neural Network (CNN) operations and nonlinearity is induced among the deep spatiotemporal representation. The CNN framework consists of various building blocks, which includes convolution, pooling and fully connected layers. Extreme Leaning Machine (ELM) is introduced for generating a feed forward framework for accurately performing the event recognition. It contains training samples, hidden layers and output matrix to generate accurate results. Experimentation result proves that proposed technique yields better accuracy and lower time complexity compared to the existing techniques.

Keywords: Video Surveillance, Video Event Recognition, Vector of Locally Aggregated Descriptors (VLAD), Convolutional Neural Network (CNN), Modified Branch and Bound Method (MBBM) and Extreme Machine Leaning (EML).