Autonomous Drone Thermal Surveillance: Edge-Optimized YOLOv8 Detection, ByteTrack Kinematics, and Heuristic Threat Classification for Border Security
Abstract
Continuous perimeter and international border monitoring presents significant operational vulnerabilities when relying on human sentries or conventional electro-optical visible sensors, particularly under zero-lux darkness, dense fog, and optical clutter. Aerial thermal long-wave infrared (LWIR) sensing mitigates optical illumination dependencies; however, uncalibrated infrared imagery suffers from thermal blooming, sensor non-uniformity, low dynamic contrast, and an absence of chromatic cues. In this paper, we propose a comprehensive, edge-deployable aerial thermal surveillance architecture developed for drone-based perimeter defense. The framework combines: (1) dynamic Contrast-Limited Adaptive Histogram Equalization (CLAHE) to amplify localized radiometric temperature gradients; (2) a lightweight YOLOv8 nano (YOLOv8n) backbone fine-tuned on FLIR thermal datasets for rapid person and vehicle detection; (3) ByteTrack association for persistent multi-object tracking across occlusions; (4) rule-based behavioral kinematics to identify anomalous postures (crouching, crawling, coordinated group motion); and (5) a dual-stage false-alarm suppression mechanism distinguishing non-human thermal emitters from genuine perimeter breaches. Benchmark evaluation demonstrates an end-to-end inference latency of 2.0 ms per frame (~500 FPS) on consumer edge GPU hardware (NVIDIA RTX 4050, AMD Ryzen 7 7000 Series), enabling fully autonomous, real-time local execution without cloud reliance. The complete system is deployed via an interactive Streamlit and TypeScript surveillance terminal. This work was awarded First Prize at the Hacker's Cult NSUT Hackathon (AI-3 Track: Computer Vision / Security Tech), Netaji Subhas University of Technology, New Delhi.
Related
- Traffic surveillance video-based a unified road accident detection, classification, and vehicle tracking framework using an optimized spatio-temporal attention mechanism
- A deep-learning based gait classification and anomaly detection framework for healthcare surveillance
- The assembly of bulge-dominated galaxies: two evolutionary channels traced through morphology, kinematics, and environment with a hybrid classification pipeline
- A multi-objective bayesian optimized neural network approach for intrusion detection
- An optimized convolutional dense network for skin cancer detection in dermatological diagnostics
BibTeX
@misc{singh2026autonomous,
title = {{Autonomous Drone Thermal Surveillance: Edge-Optimized YOLOv8 Detection, ByteTrack Kinematics, and Heuristic Threat Classification for Border Security}},
author = {Sidak Singh},
howpublished = {Zenodo},
year = {2026},
month = oct,
doi = {10.5281/zenodo.23423684},
url = {https://zenodo.org/records/23423684},
note = {CC BY-NC 3.0},
}
RIS
TY - UNPB
TI - Autonomous Drone Thermal Surveillance: Edge-Optimized YOLOv8 Detection, ByteTrack Kinematics, and Heuristic Threat Classification for Border Security
AU - Sidak Singh
PY - 2026
DA - 2026/10/16
PB - Zenodo
DO - 10.5281/zenodo.23423684
UR - https://zenodo.org/records/23423684
AB - Continuous perimeter and international border monitoring presents significant operational vulnerabilities when relying on human sentries or conventional electro-optical visible sensors, particularly under zero-lux darkness, dense fog, and optical clutter. Aerial thermal long-wave infrared (LWIR) sensing mitigates optical illumination dependencies; however, uncalibrated infrared imagery suffers from thermal blooming, sensor non-uniformity, low dynamic contrast, and an absence of chromatic cues. In this paper, we propose a comprehensive, edge-deployable aerial thermal surveillance architecture developed for drone-based perimeter defense. The framework combines: (1) dynamic Contrast-Limited Adaptive Histogram Equalization (CLAHE) to amplify localized radiometric temperature gradients; (2) a lightweight YOLOv8 nano (YOLOv8n) backbone fine-tuned on FLIR thermal datasets for rapid person and vehicle detection; (3) ByteTrack association for persistent multi-object tracking across occlusions; (4) rule-based behavioral kinematics to identify anomalous postures (crouching, crawling, coordinated group motion); and (5) a dual-stage false-alarm suppression mechanism distinguishing non-human thermal emitters from genuine perimeter breaches. Benchmark evaluation demonstrates an end-to-end inference latency of 2.0 ms per frame (~500 FPS) on consumer edge GPU hardware (NVIDIA RTX 4050, AMD Ryzen 7 7000 Series), enabling fully autonomous, real-time local execution without cloud reliance. The complete system is deployed via an interactive Streamlit and TypeScript surveillance terminal. This work was awarded First Prize at the Hacker's Cult NSUT Hackathon (AI-3 Track: Computer Vision / Security Tech), Netaji Subhas University of Technology, New Delhi.
ER -