Railway and vehicle monitoring with computer vision

THE CONTEXT

In the forestry logistics chain of southern Chile, railway crossings and industrial access points are critical: trains carrying timber cars and a high volume of trucks pass through them. Before this project, knowing what happened at each crossing —which train passed, in which direction, with how many cars and whether they were loaded or empty, which trucks entered and left— relied on manual records, phone calls, or reviewing hours of footage after the fact. There was no objective traceability nor timely information for operations and auditing.

THE DEPLOYED SOLUTION

We deployed a computer vision platform that operates autonomously and continuously (24/7) on the site's existing camera infrastructure, turning every camera into a business sensor:

  • Automatic detection of trains and trucks on live video, with object tracking across frames and automatic calculation of travel direction (entry/exit).
  • Audiovisual evidence for every event: video from the site's cameras (side and top views), plus an annotated video with detections drawn frame by frame —classification, tracking ID and confidence— ready for quick review from the web.
  • Automatic locomotive identification through industrial OCR on the registration number painted on the unit, with a rescue algorithm that corrects typical reading confusions and evidence photos for human verification.
  • Forestry load quantification: an AI model tracks each car individually and determines whether it is loaded or empty, weighting confidence frame by frame; locomotive-only maneuvers are automatically discarded. For trucks: classification by type and load status.

FORESTRY TRUCK CONTROL

The same monitoring point audits the flow of forestry trucks crossing the access: the platform detects each truck, tracks it across frames and automatically classifies its load status —loaded or empty— along with the vehicle type and travel direction.

  • Automatic logging of entries and exits with timestamp, event duration and load status.
  • Verifiable loaded/empty classification: every event is backed by annotated video with detections drawn frame by frame and evidence photos.
  • Full-cycle traceability: the same truck is recorded entering loaded and leaving empty, allowing trips to be reconciled against waybills and weighbridge records.

DETECTION CATEGORIES

The platform automatically classifies every object crossing the monitored point:

  • Trains: composition by loaded and empty cars, with locomotive identification via OCR.
  • Forestry trucks: load status (loaded/empty) and vehicle type.
  • Light vehicles: cars and pickup trucks.
  • People in the crossing area.

The taxonomy is extensible: new categories can be added on the same infrastructure according to business rules —for example fuel loading detection, heavy machinery (excavators, wheel loaders), personal protective equipment usage or operation-specific objects— by retraining the model with our labeling and data curation pipeline.

THE PLATFORM

  • Operational view with daily indicators: trains, entries/exits, forestry trucks, active locations and the latest detections with evidence.
  • Cameras grouped by location with real-time online/offline status and a georeferenced map of monitored points.
  • Interactive dashboards: activity by date, hourly distribution, average event duration, forestry composition, ranking of detected locomotives, OCR reading efficiency and heatmaps (hour vs. date) for traffic pattern analysis.
  • Email notifications for relevant events (detections, camera outages), configurable per user.
  • Full administration panel to review, correct and audit every record.

GENERATED EXCEL REPORTS

The platform produces multi-sheet Excel reports with embedded charts, ready for management without further processing: KPI summary, daily activity, detected locomotives, event duration, hourly distribution, forestry composition, per-location heatmaps and OCR reading efficiency.

CONTINUOUS MODEL IMPROVEMENT

The model is trained and improved through our own data curation pipeline: extraction of high-quality frames from real event videos, automatic removal of redundant images, model-assisted pre-labeling —humans correct instead of labeling from scratch— and deliberate inclusion of night scenes to correct bias. Each event stores an immutable copy of its detection configuration, so historical reprocessing is always consistent.

RESULTS

  • Autonomous 24/7 monitoring of railway crossings and access points, without permanent human surveillance.
  • Complete traceability: every train and truck is recorded with audiovisual evidence, direction, duration and locomotive identification.
  • Logistics visibility of the forestry chain: load composition per event, per location and over time.
  • Executive reporting ready for auditing and decision-making.
  • Scalable platform: adding a new monitoring point is a configuration process, not a development project.

← Back to success stories I want something similar