An Effective Technique for Video Condensation and Retrieval Using Convolutional Neural Network with YOLO-Aided Anomaly Detection Framework

Multiple surveillance cameras have escalated the challenge of video forensics; necessitate the compression of vast amounts of security footage to swiftly pinpoint criminal activities. With the proliferation of multimedia content including audio, photos, and videos, there’s a pressing need for effective and efficient methods of annotation and retrieval. The exponential growth in online multimedia content has led to the widespread availability of nearly identical videos that have undergone postproduction, raising significant concerns about copyright infringement and the cluttering of search results with redundant content. Moreover, the task of anomaly detection, which involves distinguishing irregularities from normal patterns, represents a basic level of video understanding. In this paper, a new deep learning-based model is developed to solve the anomaly detection difficulties in the video condensation and retrieval process. The necessary videos and their frames are gathered from online resources. Hereafter, the background extraction is conducted on the gathered video frames. These extracted feature frames are given to the You Only Look Once (YOLO) model for detecting the anomalies in the frame. The video condensation process is performed after the anomaly detection process based on the activity and time of the object. Furthermore, the required features are extracted from the condensed video using the Convolution Neural Network (CNN). Finally, the Jaccard coefficient is used to find the similarity between the extracted features from the condensed video and the query video. Lastly, the experimental analysis is performed on the developed model in terms of various metrics to verify its effectiveness.