SARD

Search and Rescue Human Detection

A computer vision project using YOLO26 for real-time human detection in search and rescue scenarios. Trained and compared two model variants (Nano and Medium), with the best model achieving 97.24% mAP50 and 72.46% mAP50-95 on the test set.

Note: This project uses the Ultralytics framework with yolo26n.pt and yolo26m.pt pretrained weights.

---

Problem Statement

In search and rescue operations, rapid identification of humans in diverse environments (aerial footage, disaster zones, wilderness) is critical. Manual review of imagery is time-consuming and prone to errors under pressure. This project develops a deep learning solution to:

1. Detect humans accurately in challenging conditions (varied lighting, occlusion, terrain)
2. Provide real-time inference suitable for drone/UAV deployment
3. Minimize false negatives to ensure no person is missed

---

Dataset

Source: SARD Computer Vision Dataset from Kaggle

Size: 5,755 images with bounding box annotations

Split Distribution:

• Training: 4,041 images (70.2%)
• Validation: 1,144 images (19.9%)
• Test: 570 images (9.9%)

Annotation Format: YOLO format (class_id, center_x, center_y, width, height - normalized)

Key Characteristics:

• Single class: human
• Diverse environments: Aerial, forest, unpaved Roads, rocks etc.
• Challenging conditions: Occlusion, poor contrast, small objects etc.
• Image resolution: 640 * 640

---

Methodology

1. Data Preprocessing & Exploration

• Label Format Conversion: Parsed YOLO format annotations (normalized coordinates) to pixel coordinates for visualization
• Data Verification: Validated all 5,755 image-label pairs for completeness and format consistency
• Exploratory Data Analysis:
  • Analyzed bounding box size distribution
  • Examined spatial distribution of humans in frames
  • Identified class imbalance (N/A - single class)
• Helper Functions: Built utilities for visualization, bbox drawing, and batch inspection

Bounding Box Distribution:

<img src="outputs/EDA_bbox_distribution.png" width="350">

Summary Statistics:

<img src="outputs/summary_statistics.png">

Sample image with bounding box:

<img src="outputs/sample_image_bounding_box.png">

2. Model Architecture

Framework: YOLO26 (Ultralytics)

Why YOLO26?

• State-of-the-art single-stage detector balancing speed and accuracy
• Optimized for real-time inference (critical for SAR drones)
• Strong performance on small objects (humans viewed from aerial perspectives)
• Built-in augmentation pipeline for robustness
• Easy deployment via ONNX export

Model Variants Trained:

• YOLO26 Nano (yolo26n.pt) - Lightweight, faster inference
• YOLO26 Medium (yolo26m.pt) - Better accuracy, moderate speed (~8.5x more parameter than Nano)

Input Resolution: 640×640

Architecture Highlights:

• Enhanced backbone for feature extraction
• Multi-scale feature fusion via PANet-style neck
• Decoupled head for classification and localization
• Optimized anchor-free detection

3. Training Configuration

Hyperparameters:

• Epochs: 300 (with early stopping, patience=15)
• Actual Training Duration: Stopped at epoch 243 (early stopping triggered)
• Batch Size: 32
• Optimizer: AdamW
• Initial Learning Rate: 1e-3
• Learning Rate Schedule: Cosine decay
• Weight Decay: 5e-4
• Image Size: 640×640
• Device: NVIDIA L40S (GPU)
• Workers: 4
• Cache Mode: RAM (for faster data loading)

Training Strategy:

• Transfer learning from pretrained YOLO26 weights
• Early stopping based on validation mAP50-95 (patience=15 epochs)
• Model checkpointing (best and last weights saved)
• Cosine learning rate decay for smooth convergence

4. Evaluation Metrics

Primary Metrics:

• Precision: When the model predicts "human", how often is it correct?
• Recall: Out of all actual humans, how many did the model detect?
• mAP50: Mean Average Precision at IoU threshold 0.5 (standard detection benchmark)
• mAP50-95: Mean Average Precision averaged across IoU thresholds 0.5 to 0.95 (stricter metric)

Loss Components:

• Box Loss: Bounding box coordinate regression error
• Class Loss: Classification error (human vs. background)
• DFL Loss: Distribution Focal Loss for precise box boundaries

---

Results

Model Performance Comparison

YOLO26 Nano:

YOLO26 Medium (Best Model):

Performance Improvement (Medium vs Nano on Test Set):

• Precision: +1.33%
• Recall: +2.20%
• mAP50: +0.84%
• mAP50-95: +4.53%

Interpretation:

• YOLO26 Medium provides the best overall performance with 94.31% precision and 94.38% recall
• High recall (94.38%) ensures most humans are detected - critical for SAR where missing a person has serious consequences
• High precision (94.31%) means low false positives (~5.7%), reducing wasted effort investigating false alarms
• 97.24% mAP50 demonstrates strong localization accuracy at standard IoU threshold
• 72.46% mAP50-95 shows good performance even with strict overlap requirements
• Medium model's 4.53 point improvement in mAP50-95 indicates significantly better bounding box precision

Training Progression (Medium)

Key Observations:

• Losses steadily decreased and plateaued around epoch 150-200
• Validation metrics stabilized with minimal overfitting
• Early stopping triggered at epoch 243 (no improvement for N epochs)

Qualitative Results (Medium)

Sample Detections:

Comparison between Nano and Medium:

Observation: Both model successfully detects partially occluded humans. Medium model consistently performs better than nano.

---

Key Insights

From Model Comparison (Nano vs Medium):

1. Model Size vs Performance Trade-off: Medium model achieves 4.53% in mAP50-95 over Nano, indicating significantly better localization precision
2. Recall Improvement: Medium model's +2.2% recall gain means detecting ~2% more humans - crucial for SAR
3. Balanced Performance: Medium model achieves balanced precision (94.31%) and recall (94.38%), avoiding the precision-recall trade-off
4. Deployment Consideration: Nano model offers 96.40% mAP50 with faster inference - viable for resource-constrained drones if speed is prioritized

From Training Process:

1. Early Stopping Effectiveness: Training stopped at epoch 243/300, preventing overfitting while achieving strong generalization
2. Transfer Learning Success: Pretrained YOLO26 weights accelerated convergence and improved final performance
3. Stable Training: Cosine LR decay with AdamW optimizer enabled smooth convergence without oscillations

From Model Performance:

1. High Recall Priority: 94.38% recall ensures minimal missed detections - critical for SAR where missing a person has serious consequences
2. Low False Positive Rate: 94.31% precision means only ~5.7% false positives, reducing wasted effort investigating false alarms
3. Robust Localization: 72.46% mAP50-95 indicates bounding boxes are well-calibrated across strict IoU thresholds
4. Strong Generalization: Test metrics (97.24% mAP50) closely match validation (97.43%), confirming model isn't overfitting

Deployment Considerations:

1. Real-Time Capable: YOLO26 architecture enables 152.3 FPS on NVIDIA L40S, suitable for drone inference
2. ONNX Export: Both models successfully exported to ONNX for cross-platform deployment (edge devices, web, mobile)
3. Model Selection: Use Medium for maximum accuracy, Nano for speed-critical applications
4. Failure Modes: Struggle with extremely small objects, and tends to perform worse on direct top down view where the leg and hand may be hidden by the head.

---

Future Improvements

1. Multi-Class Detection (If Applicable)

• Extend to detect specific categories: injured persons, rescue personnel, animals
• Enable priority tagging for triage in mass casualty events

2. Temporal Consistency (for video SAR missions)

• Implement tracking (DeepSORT, ByteTrack) to associate detections across frames
• Count unique individuals rather than per-frame detections
• Reduce false positives by filtering isolated single-frame detections

3. Adversarial Robustness

• Test on extreme weather conditions (fog, heavy rain, snow)
• Augment training data with synthetic weather overlays
• Evaluate performance degradation and set realistic confidence thresholds

4. Uncertainty Quantification

• Implement Monte Carlo Dropout or ensemble methods for confidence estimation
• Flag low-confidence regions for manual review

5. Efficiency Optimization

• Prune redundant weights to reduce model size
• Benchmark on target hardware (Jetson Nano, Raspberry Pi, drone compute modules)

6. Dataset Expansion

• Collect more examples of edge cases: prone humans, partial occlusion, camouflage clothing
• Include thermal imaging data for night operations
• Add crowd scenarios for mass-casualty incident training

---

Technologies

• Python 3.12
• Deep Learning Framework: PyTorch (via Ultralytics)
• Computer Vision: OpenCV (cv2), Python Imaging Library (PIL)
• Model Architecture: YOLO26
• Data Processing: NumPy, Pandas
• Visualization: Matplotlib, Seaborn
• Deployment: ONNX Runtime