Suspicious Human Activity Recognition From Surveillance Videos Using Deep Learning

Read full text Summarise with VIRAT
Authors: Sanki Vinod, MAHESHWARAM SIDDHARTHA, Saikam ManiVardhan
Cited by: 0
References: 2

Summary

No summary yet. VIRAT can read this paper and write a plain-language summary of its aim, methods, key findings and limitations.

Summarise with VIRAT Free to try with your Science Catalogs account.

Abstract

Suspicious Human Activity Recognition (SHAR) plays a critical role in strengthening modern surveillance infrastructures by enabling early detection and mitigation of potential security threats. The rapid growth of anti-social incidents has led to widespread deployment of CCTV systems across public and private environments. However, continuous human monitoring of large-scale video streams is impractical due to the enormous data volume generated daily For example, a typical surveillance camera that records video at a resolution of 704×576 and 25 frames per second can generate nearly 20GB of data in a single day. Because of this large amount of data, manually reviewing footage becomes both time-consuming and prone to human error. These limitations highlight the need for automated systems that can quickly and accurately identify unusual or suspicious activities while pinpointing the exact frames in which they occur. To address this problem, this study introduces a deep learning–based framework designed for reliable and efficient detection of suspicious activities in video data. The proposed approach combines Convolutional Neural Networks (CNNs), a time-distributed CNN structure, and a Conv3D model. This combination allows the system to learn both spatial information from individual frames and temporal patterns that occur across sequences of frames. Several preprocessing steps are applied to improve model performance, including frame extraction, data preparation, and optimization techniques during training. Experimental results show strong performance, with the time-distributed CNN model achieving an accuracy of 90.14%, while the Conv3D model reaches 88.23%, outperforming a number of existing methods.The models were further tested using unseen datasets and real surveillance videos sourced from YouTube, demonstrating good generalization to real-world scenarios. Overall, the proposed system improves detection accuracy and operational efficiency, making it a promising solution for developing smarter and more scalable surveillance systems that support public safety.

Fields of Study & Subjects

Human Pose and Action Recognition Anomaly Detection Techniques and Applications Video Surveillance and Tracking Methods

Publication details

Journal / source
International Journal of Engineering Development and Research
Volume
14
Issue
1
Publication date
1 March 2026
Language
EN