Long-Range and Night-Time Drone Surveillance

Project description

This six-month PFE develops a drone-based vehicle-surveillance pipeline covering detection, tracking, restoration and information extraction such as license plate, brand and color. It targets long-range/high-altitude operation and night or low-visibility conditions.

At long range, vehicles and license plates occupy only a few pixels, while turbulence, haze and jitter degrade images. At night or in fog, an RGB camera becomes unreliable, so a thermal infrared camera is added. Thermal imagery introduces limitations including lower resolution, lack of color, unreadable plates and RGB/thermal misalignment.

Goals

• Develop long-range perception for vehicle detection, tracking and identification under motion blur and atmospheric turbulence, using multi-frame super-resolution optimized for recognition and abstention when evidence is insufficient.
• Use RGB-thermal fusion to track vehicles day and night, reading plates from RGB whenever possible.
• Measure performance by distance and lighting and create an operational envelope recommending actions such as descending, zooming or switching sensors when identification is unreliable.

Work to be performed

• Assemble paired RGB-thermal data from public drone datasets labeled by altitude, angle and lighting.
• Train and compare detectors, trackers and super-resolution models across distance bands.
• Test robustness to blur and turbulence.
• Build and compare RGB-only, thermal-only and fused systems.
• Build the operational-envelope model and integrate the complete pipeline into a live-overlay demonstration.

Required skills

Must-have: strong PyTorch skills, clean ML engineering practices, computer-vision experience in detection, tracking and image restoration, and understanding of camera geometry and multi-camera calibration.

Nice-to-have: thermal infrared imaging, RGB-thermal fusion, super-resolution, small-object detection and edge deployment using TensorRT or quantization.

Deliverables and schedule

Deliverables include a long-range perception module, RGB-thermal fusion module, benchmark and operational-envelope model, integrated live-overlay demo, technical report and reproducible repository. Months 1–2 cover data assembly, detection, tracking and distance estimation; months 3–4 cover super-resolution and RGB-thermal fusion; months 5–6 cover the operational envelope, integration, final report and handover.

Monthly stipend with potential end-of-internship performance bonus and potential paper-publication co-authorship.