Author: Prathamesh Chatorikar
Affiliation: University of California, Santa Cruz
Date: March 2025
This repository introduces a hybrid system for identifying safe landing zones for drones using aerial imagery. The system leverages state-of-the-art object detection (YOLOv11) in conjunction with spatial reasoning powered by large language models (GPT-4 Turbo). The methodology ensures optimal landing zone selection by avoiding hazardous regions such as fire, smoke, debris, and solar panels, and by prioritizing zones such as rooftops and ships when available.
The framework operates in two stages:
-
YOLO-Based Detection
Utilizes YOLOv11 Nano to detect both hazardous and safe zones in aerial imagery. -
Post-Processing and Reasoning
Applies a custom-designed algorithm (Algorithm 1) to evaluate rooftop safety. In cases where no safe zones are detected or validated, a fallback mechanism invokes GPT-4 Turbo to reason spatially about the best landing location using either:- Object-level metadata (Hazard-aware mode), or
- Entire image context (Full-image mode)
Author: Prathamesh Chatorikar
- Model Architecture: YOLOv8 Nano (Ultralytics)
- Training Dataset:
Custom Roboflow dataset with the following class labels:- Safe:
rooftop,ship - Hazardous:
fire,smoke,debris,solar-panels
- Safe:
- Labeling: Manual annotation with geometric augmentations
- Training Infrastructure: NVIDIA T4 GPU via Google Colab
- Output Artifacts: Best weights saved as
best.pt
Comprehensive timing analysis is conducted across three operational modes:
- YOLO inference time
- LLM-based reasoning with hazard metadata
- LLM-based reasoning using full-image prompts
Results are visualized to compare efficiency and latency across modes.
A complete report including background, implementation details, dataset construction, evaluation metrics, and visual results is provided below.
- Python 3.8+
- OpenAI API access
- Ultralytics YOLOv11 installed