Inspiration
When severe storms strike, electrical infrastructure can suffer widespread damage, leaving thousands of people without power. Utility companies like CenterPoint Energy must inspect extensive networks of power lines, poles, transformers, and substations to identify hazards and prioritize restoration efforts. Traditional inspections can require personnel to review hours of drone footage or physically visit potentially dangerous locations. We wanted to explore how artificial intelligence could accelerate this process, reduce manual review, and help crews identify the most urgent hazards. For the CenterPoint Energy × UH Code Coogs Hackathon 2026 — Track 1: Drone Imaging, we developed Infinity Vision, an AI-powered system designed to turn drone and vehicle imagery into actionable infrastructure inspection findings.
What it does
Infinity Vision analyzes uploaded photographs and videos to identify potential hazards affecting electrical infrastructure. Our application follows a three-stage workflow:
- Identify hazards Using AI-powered object detection and a custom rules engine, Infinity Vision searches for potential hazards, including:
- Flooded electrical infrastructure
- Downed power lines and utility poles
- Fallen trees near or potentially contacting electrical equipment
- Leaning utility poles
- Flooded access routes
Fallen vegetation
Record and prioritize findings Each detected hazard becomes a finding containing its category, detection score, timestamp, supporting image, AI-suggested severity, and GPS information when available. Our custom severity system categorizes potential hazards as CRITICAL, HIGH, MEDIUM, or LOW, allowing inspectors to prioritize urgent findings. These severity levels are heuristic recommendations rather than validated engineering risk assessments.
Review and generate work orders Infinity Vision provides an interactive dashboard where inspectors can review findings, confirm hazards, mark false positives, request on-site inspections, and adjust severity levels. Confirmed high-priority findings generate proposed work orders identifying the appropriate response crew. Findings with location data can also be displayed on a map.
How we built it
We developed Infinity Vision primarily in Python, combining a pretrained computer vision model with custom image-processing algorithms, hazard-detection rules, and an interactive web application.
AI Object Detection — Grounding DINO and Hugging Face We used Grounding DINO, an open-vocabulary object detection model, to identify utility poles, power lines, transformers, substations, trees, and floodwater. Unlike traditional classifiers, Grounding DINO can locate multiple objects within an image using text prompts. This allowed us to build a prototype without training a custom model from scratch.
Hazard Analysis — Custom Python Rules Engine We developed a rules engine that evaluates detected objects, their positions, and their relationships to identify potential hazards. The system also accounts for camera perspective, distinguishing street-level, angled aerial, and top-down drone footage to reduce incorrect hazard classifications.
Pole Tilt Detection — OpenCV Using Canny edge detection and Hough line fitting, we implemented an algorithm to estimate utility pole tilt from street-level imagery. We also developed camera-tilt correction logic to reduce false alarms caused by rotated cameras.
Image and Video Processing — OpenCV and Pillow OpenCV processes videos and extracts frames for analysis, while Pillow reads image metadata, including GPS coordinates when available. We also implemented support for DJI metadata and subtitle-based flight logs to associate findings with drone capture locations.
Interactive Dashboard — Streamlit We built a Streamlit interface that allows users to upload imagery, run AI analysis, review prioritized findings, inspect supporting images, and manage proposed work orders.
Inspector Feedback — Learning Memory Infinity Vision records inspector decisions and uses visual fingerprints and learned confidence thresholds to help prioritize future findings. This feedback mechanism operates separately from Grounding DINO; it does not retrain the underlying model.
Challenges we ran into
One of our biggest challenges was the lack of labeled datasets specifically designed for detecting electrical infrastructure hazards. Because we had less than 24 hours to develop our prototype, training a custom model from scratch was impractical. We initially experimented with YOLO-World, but it produced no detections on our test imagery. We subsequently switched to Grounding DINO, which performed better for our initial use cases. Another significant challenge was distinguishing actual hazards from visually similar objects. For example, the model sometimes confused substation structures with fallen trees or identified equipment sheds as transformers. We also encountered false flooding alerts in safe drone footage. To address these limitations, we developed additional rules incorporating camera perspective, vegetation-color analysis, and inspector verification. We also discovered that aerial imagery introduces unique challenges. A pole that appears tilted may actually be vertical, and a tree appearing to overlap a power line may not physically touch it. These challenges reinforced the importance of combining AI predictions with contextual analysis and human oversight.
Accomplishments that we're proud of
Within a single hackathon, our team developed a working prototype that connects computer vision with an infrastructure inspection workflow. Some of our key accomplishments include:
- Integrating Grounding DINO for zero-shot object detection without requiring a custom-trained model.
- Building a system that processes both photographs and drone videos.
- Implementing a custom hazard classification and severity prioritization system.
- Developing pole-tilt estimation with camera-rotation correction.
- Reducing false alarms in a routine top-down drone test video from 26 to 2 after implementing camera-aware rules.
- Incorporating GPS metadata and DJI flight-log support.
- Creating an interactive review dashboard with human verification, proposed work orders, and inspection tracking.
- Implementing an inspector-feedback memory and training-data export workflow.
- We also designed the model to run locally after its initial download, allowing inference without sending imagery to an external AI API.
What we learned
This project taught us that detecting an object and understanding whether it represents a hazard are two different problems. A computer vision model may identify a tree, power line, or transformer, but determining whether that object is damaged or dangerous requires additional context. We gained hands-on experience with pretrained vision models, object detection, video frame processing, image metadata, geometric analysis, and AI-assisted decision-making. We also learned the importance of evaluating false positives, accounting for camera perspective, and designing systems where human expertise remains central to safety-critical decisions.
What's next for Infinity Vision
Our long-term vision is to develop Infinity Vision into a scalable infrastructure monitoring platform.
Future improvements include:
- Custom model training: Use inspector-reviewed imagery to build a labeled dataset and train a specialized utility-hazard detector.
- Improved accuracy: Evaluate precision and recall on a larger, manually labeled dataset and refine hazard-detection thresholds.
- Real-time analysis: Optimize inference for faster processing and potential deployment on drones or utility vehicles.
- More accurate geolocation: Combine drone position, camera orientation, and geographic data to estimate the actual location of damaged infrastructure.
- Enterprise integration: Connect findings with utility GIS platforms, outage-management systems, and existing work-order workflows.
- Improved hazard assessment: Develop more reliable methods for distinguishing physical infrastructure damage from misleading visual overlap. Our goal is to help utility companies move from hours of manual footage review toward faster, more organized, and safer infrastructure inspections.
