Optimizing the deep features using the spectral concept for an efficient video classifications

The identification of suspicious activity is greatly influenced by the efficient learning and classification of video data. Learning large videos requires more processing time than conventional deep learning methods, because deep features have relatively large dimensions. In the present investigation, a spectral-based deep learning method for producing reduced or lower dimensional representations of surveillance videos is developed. There are two essential steps in the proposed method. In order to identify suspicious activity, the first method involves transforming the deep characteristics of video data into lower-manifold spectral space using classification techniques. The lower-rank representation of the video data is found by computing a Laplacian matrix. To show the effectiveness of the suggested strategy in comparison to current methods, experiments are conducted on benchmarked video surveillance datasets.