Summarizing Surveillance Videos through Automatic Human Activity Detection

In today’s world, ensuring the safety of individuals relies heavily on video surveillance encompassing activity monitoring, recording, and event detection. However, this surveillance approach introduces challenges such as continuous 24 × 7 monitoring, substantial storage requirements, and the manual identification of objects or events of interest within recorded video sequences. This paper addresses these practical issues by presenting a video summarization and automatic person identification technique based on deep learning. The objective is to generate concise video summaries from surveillance footage that are shorter than the original video length and contain relevant keyframes featuring both moving and still persons. The proposed technique integrates the YOLOv5m and DeepSORT frameworks for detecting and tracking human beings, facilitating the creation of these short video summaries. Experimental evaluations are conducted on the VIRAT and TVSum video datasets, and the hybrid approach is validated using real-life video recordings from institutional premises and public spaces. The results indicate that the proposed method automatically and effectively detects persons under varying lighting conditions, thereby producing informative summary videos.