Model Card: YOLOv5 Wheelchair Detection Model

Model Details

Basic Information

• Model Name: YOLOv5 Wheelchair Experiment (High Accuracy Configuration)
• Model Type: Object Detection
• Architecture: YOLOv5
• Framework: PyTorch/Ultralytics
• Version: Custom trained variant
• Date: 2025
• License: [Specify license]

Model Architecture

• Layers: 157
• Parameters: 7,225,885 (~7.2M)
• Gradients: 0 (inference mode)
• Computational Complexity: 16.4 GFLOPs

Intended Use

Primary Use Cases

• Multi-class object detection across 80 COCO dataset categories
• Accessibility and mobility-related detection tasks
• Real-time object recognition in images and video streams
• Wheelchair and pedestrian environment analysis

Intended Users

• Accessibility technology developers
• Computer vision researchers
• Robotics and autonomous systems engineers
• Urban planning and infrastructure analysts

Out-of-Scope Uses

• Medical diagnosis or clinical decision-making
• Surveillance applications without proper consent
• Any application requiring 100% reliability for safety-critical decisions

Training Data

Dataset

• Source: COCO (Common Objects in Context) dataset classes
• Classes: 80 object categories including people, vehicles, animals, household items, and outdoor objects
• Evaluation Set: 128 images, 929 object instances

Data Preprocessing

• Image resizing and normalization
• Data augmentation (standard YOLO pipeline)

Performance

Overall Metrics (Validation Set)

Top Performing Classes

• Motorcycle: 92.5% P, 100% R, 99.5% mAP@50, 96.8% mAP@50-95
• Zebra: 90.1% P, 100% R, 99.5% mAP@50, 99.5% mAP@50-95
• Bear: 77.0% P, 100% R, 99.5% mAP@50, 99.5% mAP@50-95
• Train: 86.7% P, 100% R, 99.5% mAP@50, 96.3% mAP@50-95
• Donut: 96.1% P, 100% R, 99.5% mAP@50, 92.6% mAP@50-95

Challenging Classes (Lower Performance)

• Car: 81.8% P, 39.1% R, 53.4% mAP@50, 29.9% mAP@50-95
• Baseball glove: 87.5% P, 42.9% R, 43.0% mAP@50, 27.4% mAP@50-95
• Traffic light: 73.5% P, 42.9% R, 56.3% mAP@50, 30.5% mAP@50-95
• Book: 80.8% P, 51.7% R, 67.8% mAP@50, 42.2% mAP@50-95
• Cell phone: 59.2% P, 75.0% R, 71.0% mAP@50, 45.2% mAP@50-95

Performance Characteristics

• Strengths: Large animals, vehicles (motorcycles, trains), distinctive food items, furniture
• Weaknesses: Small objects, highly variable objects (cars in different contexts), items with subtle features (books, handbags)

Limitations

Technical Limitations

• Lower recall on common urban objects (cars at 39.1%)
• Struggles with small objects and dense scenes
• Performance varies significantly across object classes
• May not generalize well to domains outside training distribution

Ethical Considerations

• Model trained on COCO dataset may inherit biases present in that dataset
• Person detection should be used responsibly with privacy considerations
• Performance may vary across different demographic groups or geographic regions
• Should not be used as sole decision-maker in safety-critical applications

Known Issues

• Reduced performance on vehicles in varied contexts
• Traffic light detection may be unreliable (42.9% recall)
• Book and handbag detection shows high miss rates

Recommendations

Best Practices

• Use confidence thresholding appropriate to your application
• Combine with temporal filtering for video applications
• Validate performance on your specific domain before deployment
• Consider ensemble methods for critical applications
• Regularly monitor performance in production

Hardware Requirements

• Minimum: CUDA-capable GPU with 4GB+ VRAM for real-time inference
• Recommended: Modern GPU (RTX 3060 or equivalent) for optimal performance
• CPU inference possible but significantly slower

Inference Speed

• Real-time capable on modern GPUs (30+ FPS typical)
• Speed depends on input resolution and hardware

Model Card Authors

Model Card Contact

Citation

References

• Ultralytics YOLOv5: https://github.com/ultralytics/yolov5
• COCO Dataset: https://cocodataset.org/
• Model results saved to: wheelchair_runs/wheelchair_exp_high_accuracy

---