Downloads: 1
Student Project | Computer Science and Engineering | Volume 15 Issue 4, April 2026 | Pages: 491 - 498 | India
Deep Learning Based Suspicious Activity Detection in Surveillance Systems
Abstract: With the increasing presence of CCTV installations in modern smart cities, there is now a need for more advanced video analytics techniques to automatically detect threats. Traditional surveillance frameworks treat the detection of spatial objects and the identification of temporal activities as two independent phases, resulting in a reduced ability to provide contextually rich insights and increased numbers of false alarms. In this work, we present a novel deep learning framework where both object localisation and activity classification are conducted simultaneously using one single inference graph. We utilise the YOLOv8 model as the spatial encoder, capable of detecting people and suspicious objects with pixel-accurate boundary annotations, while an auxiliary dual stream network is used to detect activities like fighting, loitering, and aggressive gestures from the optical flow and RGB visual streams. A multimodal feature fusion mechanism is applied to combine the features extracted from spatial object detection with those of the temporal activities, followed by a rule-based multi-layer decision making engine to compute risk scores based on a configurable alert threshold. Using experiments conducted on the Roboflow Dangerous Action Detection dataset, we achieve a 91.3% precision, 88.6% recall, and an F1 score of 89.9% with an average computational latency of 28ms per frame. Our method offers a 32.9% increase in accuracy over traditional human-in-the-loop surveillance systems, and our solution decreases the occurrence of unnecessary alerts by 18.7%.
Keywords: YOLOv8, Deep Learning, Suspicious Activity Detection, Behaviour Analysis, Feature Fusion, Real-Time Surveillance, Object Detection
How to Cite?: Pallela Anurag, Gaddam Esha, Doddi Sri Varsha, Deepthi Joshi, "Deep Learning Based Suspicious Activity Detection in Surveillance Systems", Volume 15 Issue 4, April 2026, International Journal of Science and Research (IJSR), Pages: 491-498, https://www.ijsr.net/getabstract.php?paperid=SR26405080639, DOI: https://dx.dx.doi.org/10.21275/SR26405080639