How memetsumer/airsim-uav-sar-yolov5 works
Project type and purpose
• This is a research-oriented simulation pipeline rather than a production service‐oriented application.
• Purpose: run an autonomous UAV in the AirSim simulator, capture imagery, and perform Search‐And‐Rescue (SAR) object detection using different YOLOv5 model variants (nano, small, medium).
File structure
• Top‐level files:
• LICENSE (project license)
• README.md (project description)
• Three Jupyter notebooks, one per YOLOv5 variant:
– YOLOv5n_AirSim_SAR_UAV_17_05_2022.ipynb
– YOLOv5s_AirSim_SAR_UAV_17_05_2022.ipynb
– YOLOv5m_AirSim_SAR_UAV_17_05_2022.ipynb
Readme insights
• Confirms the workflow: AirSim provides simulated UAV flight and camera feed; YOLOv5 models run on that stream; predictions are logged for analysis.
Components
a. Main components
1. AirSim Simulator
2. UAV Agent (PX4 or built-in AirSim drone API)
3. Camera/Sensor Module (virtual RGB/Depth/LiDAR)
4. Inference Engine (YOLOv5 variants)
5. Prediction Logger (predictions_yolov5 directory)
b. Relationships and interactions
– Control flow:
AirSim Simulator → UAV Agent → captures frames
Frames → Inference Engine → produces detection results
Results → Prediction Logger → stored artifacts
c. Architectural patterns / design principles
– Pipeline architecture: modular data acquisition, modular inference, modular logging
– Experiment-driven: interchangeable model components (n, s, m)
– Clear separation of concerns: simulation vs. perception vs. logging
d. Relevant technologies / libraries
– AirSim (Microsoft AirSim Python/C++ API)
– YOLOv5 (PyTorch‐based detection code)
– Jupyter Notebooks for orchestration and visualization
– Python logging/File I/O for predictions
– (If used) camera API, NumPy/OpenCV for frame handling