Skip to main content
    Computer Vision Pipeline for Drone Infrastructure Inspection
    Energy & Industrial Infrastructure

    Computer Vision Pipeline for Drone Infrastructure Inspection

    ML pipeline that turns drone footage of energy and industrial assets into geotagged, inspector-confirmed defect findings and draft inspection reports.

    Last updated 2026-11-19

    Client details are confidential under NDA. Figures are approximate and rounded.

    Client
    Infrastructure inspection company
    Industry
    Energy & Industrial Infrastructure

    Project Overview

    An infrastructure inspection company flies drones over energy and industrial assets and delivers condition reports to asset owners. Its inspectors were reviewing hours of footage per site by hand to find a small number of defects. TekPiq built a computer-vision pipeline that ingests the footage and flight telemetry, detects and segments likely defects, routes them to inspectors for confirmation, and places confirmed findings on a map and in the asset register.

    Key Challenges

    Large batches of video and high-resolution stills per site, much of it overlapping or showing no defect at all

    Defects that are small, rare and visually similar to harmless features such as stains, shadows and reflections

    Limited labeled training data at the start, and inspectors with little time to label more

    Findings that are only useful if they can be tied to a specific asset and location, not just a frame number

    A requirement that inspectors, not the model, make the final call on every reported defect

    Deciding which workloads belong on cloud GPUs and which could run closer to the drone

    Our Approach

    We started by fixing an evaluation set: a held-out collection of labeled frames, reviewed by the client's inspectors, against which every model version is scored for precision and recall per defect class. Ingestion pulls video, stills and flight telemetry (GPS position, altitude, heading) into a single batch per inspection. Frames are sampled according to flight speed and altitude, then near-duplicates are removed so the models see distinct views of each asset. YOLO-family detectors and segmentation networks, trained in PyTorch and exported to ONNX, flag and outline candidate defects. Detections go to a reviewer interface where inspectors confirm, reject or correct them; those decisions feed an active-learning loop that prioritises the most informative frames for the next labeling round. Confirmed findings are geotagged from the telemetry, matched to assets in the register and assembled into a draft report. Production inference runs as cloud GPU batch jobs after each flight; ONNX export keeps an edge deployment path open for cases where results are needed on site.

    Technologies Used

    Python & PyTorch
    YOLO-family detection & segmentation networks
    ONNX model export
    Cloud GPU batch inference
    Flight telemetry & geospatial processing (GPS, altitude)
    Active-learning labeling workflow
    Web-based reviewer interface

    Related services

    Key Features

    Batch ingestion of drone video, still images and flight telemetry per inspection

    Frame sampling and near-duplicate removal before inference

    Defect detection and segmentation models with per-class confidence thresholds

    Reviewer interface where inspectors confirm, reject or correct each detection

    Active-learning loop that turns reviewer decisions into new training labels

    Geotagging of confirmed findings onto a map and the client's asset register

    Draft inspection report generation from confirmed findings

    Evaluation set with precision and recall tracked per defect class for every model release

    Cloud GPU batch processing, with ONNX export as the path to edge inference

    Business Outcomes

    Inspectors review flagged frames instead of watching full footage

    Every reported defect is confirmed by an inspector before it reaches a client report

    Findings arrive geotagged and linked to a specific asset, not just a timestamp in a video

    Each model release is compared with the previous one on the same evaluation set, per defect class, before deployment

    Reviewer decisions become training data, so the labeled dataset grows as part of normal inspection work

    Report drafting starts from structured findings rather than from screenshots and notes

    Transformative Impact

    The pipeline changes where inspector time goes: from searching footage for defects to judging the candidates the models surface. Because every model release is scored on a fixed evaluation set and every reported finding is confirmed by a person, the client can explain to asset owners how a defect was found and who signed it off.

    Ready to Transform Your Business?

    Let's discuss how we can create a customized solution that delivers measurable results for your organization.