Computer Vision Thermal Surveillance System

Overview

A thermal-imagery surveillance stack built for 24/7 operation in the field. Deep-learning detectors run on edge hardware co-located with the thermal camera, producing real-time alerts without relying on a backhaul link. The pipeline is tuned for the quirks of thermal input — low texture, high noise at range, temperature drift — and hardened for rugged deployment.

The Problem

RGB surveillance fails at night and in degraded visibility. Thermal gives you a signal in those conditions but doesn't look like the RGB data detectors are pre-trained on. Off-the-shelf models perform poorly; operators either miss detections or drown in false positives. Purpose-built thermal models, trained on in-domain data and deployed at the edge, close that gap.

My Role & Contribution

• Curated and annotated the in-domain thermal dataset and defined evaluation protocols
• Trained and iterated on YOLOv8-based detectors fine-tuned for thermal input
• Built the edge deployment pipeline, including TensorRT acceleration and the ruggedized runtime

Approach

• Domain-specific data curation — thermal imagery across range, weather, and time-of-day variations
• YOLOv8 detectors fine-tuned on thermal data with augmentations tailored to thermal noise characteristics
• TensorRT conversion and INT8/FP16 quantization for real-time inference on edge GPUs
• OpenCV-based pre/post-processing pipeline with temporal smoothing to suppress flicker false-positives
• Dockerized runtime on Linux edge boxes for reproducible deployment and updates

Tech Stack

Results & Impact

• Real-time detection on edge hardware without backhaul dependency
• Deployed in operational use for continuous monitoring