Enhancing Drone-Based Vehicle Classification: YOLO Model Adaptation to HEF Extension

Problem Statement

• Need for real-time, accurate detection of vehicle types (military vs. civilian) from drone footage.
• Standard YOLO models were too heavy and not optimized for drone hardware (resource constraints).

Solution

• Trained a YOLO-based model for vehicle classification.
• Adapted the trained YOLO model to HEF (Hardware-Efficient Format) extensions to make it lightweight and deployable on drones.

Process / Approach

• Data Collection

  Gathered aerial images of military and civilian vehicles.

• Annotation

  Properly labeled datasets for clear distinction

• Model Training

  Trained a YOLOv8 model & o Fine-tuned for high-altitude, low-resolution images

• Conversion

  Optimized the trained model to HEF using model compression and quantization techniques

• Deployment

  Deployed the HEF model onto drone hardware & Tested in various environments to ensure robustness

Key Features / Highlights

• Real-time vehicle detection from drone footage.
• Clear differentiation between military and civilian targets
• Lightweight model optimized for resource-limited devices (like drones).

Results / Impact

• Model Accuracy

  Achieved 92% detection accuracy on test datasets

• Performance

  Reduced model size by 60% after conversion to HEF without significant loss in accuracy

• Deployment

  Model successfully tested on drone simulator and real-time drone hardware.

Challenges and Learnings

• Challenge

  Low-resolution aerial images made classification harder

• Solution

  Data augmentation techniques and fine-tuning of YOLO anchors

• Challenge

  Maintaining model accuracy after compression

• Solution

  Careful hyperparameter tuning during HEF conversion

• Learning

  Need to balance model size, speed, and accuracy carefully for drone systems