Vision-Based Drone Surveillance System for Suspicious Activity Detection Using Deep Neural Networks

Authors: Sayali Dhomase, Jayshri Niphade, Pooja Bhosale, Prerana Jadhav

Country: India

Full-text Research PDF File:   View   |   Download

Abstract: In response to the growing need for intelligent and responsive surveillance in public spaces, this study presents a real-time drone-assisted system designed to detect suspicious or violent human behavior using advanced deep learning and edge computing techniques. Traditional CCTV systems often lack the capability for timely threat detection, prompting the development of a more dynamic and efficient approach. The proposed system employs a drone equipped with a high-resolution camera to capture aerial video footage, which is enhanced using a Blind Deconvolutional Algorithm for improved image clarity. Human presence is detected through the Faster R-CNN Inception V2 model, and pose estimation is carried out using MediaPipe to extract body keypoints accurately. A hybrid model combining MobileNetV2 for spatial analysis and Long Short-Term Memory (LSTM) networks for temporal behavior analysis ensures efficient recognition of suspicious activities. To support real-time performance on low-power edge devices such as the Raspberry Pi, the system is optimized with
TensorFlow Lite. Upon identifying abnormal or violent behavior, the system promptly sends alerts via Gmail or Telegram Bots to notify concerned authorities. Experimental evaluations conducted on public datasets and custom aerial footage indicate high accuracy, low latency, and enhanced operational efficiency over traditional systems. The framework is scalable and privacy-conscious, making it well-suited for implementation in environments such as school campuses, public gatherings, and critical infrastructure, thereby enhancing proactive public safety measures in smart city infrastructures.

Keywords:

Paper Id: 232550

Published On: 2025-06-03

Published In: Volume 13, Issue 3, May-June 2025

Share this