MULTI-VIEW VIDEO SUMMARIZATION USING TARGET DETECTION AND CLUSTERING

Abstract

In the era of booming technology, with the advancement of mobile phones and camera-enabled devices the application and purpose of digital data have been exponentially increased. The data collected by these devices are in trillions to larger units of digital data. Therefore, it has become quite uneasy to retrieve valuable information from these videos. Here, it is multiview video summarization, the concept of understanding and finding out the important information from a large video file when inspected through different angles and projections. In this project, we proposed deep learning techniques with a clustering algorithm in three phases for multi-view video summarization. In the first phase, shot segmentation is done using target/object-based along with eliminating redundant frames. The second phase extracts frame-level features using the ResNet50 CNN model and passes them to the final step. In the third or final phase, the visual features are clustered using HDBSCAN and select the final keyframes based on entropy value (informativeness). Experimental results on the popular datasets clearly shows that the proposed methodology performs better than the existing methods.