Video event detection, classification and retrieval using ensemble feature selection

Abstract

In recent times, video analysis gains more popularity among the researchers, due to its social impact and widespread applications. In this research article, a novel algorithm is proposed to detect the multiple events in the videos. At first, histogram equalization technique is used for contrast enhancement and smoothing of the videos, which are collected from Columbia consumer video, and UCF101 databases. The histogram equalization delivers a better quality of frames without loss of any local information like edges, patches and points. Further, feature extraction is accomplished using gradient local ternary pattern, histogram of oriented gradients, and Tamura features to extract the feature vectors from the enhanced frames that speed up the training process. The extracted feature vectors are high dimension in nature, so an ensemble feature selection algorithm is proposed to select optimal features from the extracted features. The selected optimal feature sub-sets are fed to multi support vector machine classifier to classify the multiple events. Finally, the relevant events are retrieved utilizing Euclidean distance measure and the simulation result showed that the proposed algorithm achieved better performance in multi-event recognition. The proposed algorithm almost showed a maximum of 14.21\u2009% improvement in video event classification by means of accuracy compared to the existing algorithms like two-stage neural network, multi-stream deep learning system, improved two-stream model, Concept wise Power Law Normalization (CPN) with Convolutional Neural Network (CNN), Motion based Shot Boundary Detection (MSBD) algorithm, and motion features with Recurrent Neural Network (RNN).