Vision#

Modern infrastructure validation is slow, manual, and error-prone.

This project introduces an AI-powered validation framework that:

• Detects poles
• Verifies ownership
• Assesses structure
• Provides decision-ready outputs

Project Attributes#

Post Files#

1
src/content/posts/
2
└── pole-detection/
3
    ├── cover.png
4
    └── index.mdx
5

6
GitHub Repo: dranubhaparashar/Pole-Detection/
7
├── dataset/
8
├── runs/
9
│   └── detect/
10
├── 001_png.rf.72b75e71c8e9e08119e23830becb46ea.jpg
11
├── 001_png.rf.72b75e71c8e9e08119e23830becb46ea.txt
12
├── 1.ipynb
13
├── 1.py
14
├── README.md
15
├── app.py
16
├── data.yaml
17
└── yolo26n.pt

Why This Matters#

NOTE

Manual validation involves 13+ checks across multiple systems.

IMPORTANT

Wrong validation → cost impact + safety risks + delays

TIP

AI enables confidence-based automated decisions

WARNING

Low recall indicates dataset limitation and class imbalance

CAUTION

Decisions should not be fully automated until model recall and validation coverage improve.

What This System Does#

Core Capabilities#

• Pole detection (YOLO26n)
• Ownership verification (AEP / Frontier DB)
• Structural assessment (visual + metadata)
• Space feasibility analysis
• Multi-source reconciliation
• Confidence-based decision engine

Bento Overview#

YOLO26n Detection Engine#

Inference#

1
from ultralytics import YOLO
2

3
model = YOLO("best.pt")
4

5
results = model("pole.jpg")
6

7
for r in results:
8
    print(r.boxes)

Confidence Engine#

1
def confidence_score(pole, ownership, geo, structure):
2
    return (
3
        pole * 0.3 +
4
        ownership * 0.2 +
5
        geo * 0.3 +
6
        structure * 0.2
7
    )

Decision Logic#

1
def decision(score):
2
    if score > 0.85:
3
        return "Auto-Approve"
4
    elif score > 0.55:
5
        return "Manual Review"
6
    else:
7
        return "Reject"

Example Threshold Update#

1
AUTO_APPROVE_THRESHOLD = 0.90
2
MANUAL_REVIEW_THRESHOLD = 0.60
3
AUTO_APPROVE_THRESHOLD = 0.85
4
MANUAL_REVIEW_THRESHOLD = 0.55
5
REJECT_THRESHOLD = 0.00

Runtime Example#

1
python train.py --model yolo26n.pt --data data.yaml --epochs 100 --batch 16 --imgsz 640
2
python infer.py --weights best.pt --source pole.jpg
3
python evaluate.py --predictions outputs/results.json

Performance Snapshot#

WARNING

Low recall indicates dataset limitation and class imbalance

CAUTION

This system should remain human-assisted until model recall and evidence consistency improve.

Current Challenges#

• Small dataset
• Class imbalance
• Low recall

Next Improvements#

• Increase dataset size
• Improve labeling
• Reduce class complexity
• Apply active learning

Demo Videos#

Key Innovation#

IMPORTANT

This is not just object detection —
it is Decision Intelligence for Infrastructure

Conclusion#

This project transforms:

Manual Validation → AI-Assisted Decision System

• Combines vision + data + reasoning.
• Provides explainable decisions.
• Enables faster approvals.

Final Thought#

From detecting poles → to understanding and deciding infrastructure

The final recommendation is AI-assisted, not fully autonomous.