Anomaly Detection using C3D Networks in Surveillance videos

Abstract

Anomaly detection is a field of Video Analysis. It has been attracted by diversified research fields and application areas. Anomaly events might be outlined by the alteration from the usual or normal but not naturally in an undesirable way. The major challenges of detecting anomaly includes difficulty in designing the models because their uncertainty and their nature of dependency on the situation of the scene. Now a days,Convolutional Neural Networks (CNN), a division of deep learning field have exhibited tremendous improvement in detecting anomalies from the surveillance videos. But newlineCNN s failed in maintaining the accuracy for real-time videos due to several factors. newlineSuch as noise, variations in defining various events, definite context, limited training newlinedata requirement of computing resources and lack of temporal learning. The inefficiency newlinein maintaining temporal features of the videos, the Convolutional 3D networks newlineis designed in such a way that it takes entire video as input perform 3D convolution on multiple video frames and preserves the spatial as well as temporal features over the newlinenetwork, even though C3D models giving more prominent results, still there are some newlinechallenging issues related to the time to learn the larger videos and dimensionality issues of the video data set.The research work explores two contemporary approaches to recognise anomaly events in video surveillance. Three main components decide the efficiency and accuracy of the anomaly detection method. Selection of Deep Learning model, Preprocessing methods and Classification model.The first proposed approach is selection of C3D CNN for spatial and temporal feature extraction. The second proposed approach used hybrid Kalman-IPCA filter in preprocessing newlinefor extracting high ranked features by eliminating noise and pretrained C3D newlineCNN along with Ensemble classifier used for one class classification approach. Due to newlinethis the accuracy and the time to train the network are improved. The third proposed newlineapproach is a Hybrid Ensemble feature ext