Action Recognition Using a Deep Neural Network for Video Surveillance

¹ All India Shri Shivaji Memorial Society Institute of Information Technology, Pune (SPPU university), India
² Bharati Vidyapeeth’s College of Engineering for Women, Pune (SPPU university), India

^* Corresponding author: thenged@gmail.com

Published online: 18 June 2025

Abstract

These days, video surveillance systems are employed in a wide range of settings, both public and private. Identifying human interactions is one of the most problematic components of video surveillance. At each node, these surveillance systems can accurately identify persons or events. When it comes to video surveillance, suspicious actions in the monitoring stream are unusual. As a consequence, human monitoring of suspicious actions may become rather exhausting, compromising dependability and timeliness during times of urgency due to monitoring fatigue. This emphasises the vital need of identifying suspicious behaviour. Most of these systems employ deep learning and segmentation approaches to achieve outstanding performance. However, these algorithms have a high computational cost, limiting their capacity to operate in real time. To accomplish this objective, we introduce a novel deep learning system with a single stage that can be taught from start to end. This system can optimise spatial and temporal activity classification. When employing the UCF crime dataset, the recommended architecture achieves 96% accuracy in its analysis. This is in contrast to the results of cutting- edge algorithms (SOTA).