Railway Container Number Automatic Recognition System Solution

To improve logistics efficiency and data accuracy, this solution captures container footage via cameras and identifies container numbers with advanced computer vision technology from Chushi. It covers three application scenarios: incoming railway trains, reach stacker handling operations, and semi-trailers hauling containers exiting the yard.

1. Camera Selection

High-resolution IP cameras with night vision capability are deployed to capture crisp images under varying lighting and weather conditions.

2. Access Modes

-Wired Access: Cameras connect to local area network via Ethernet cables for stable and high-speed data transmission.

-Wireless Access: WiFi or 4G-enabled cameras are adopted where cabling is impractical to guarantee real-time data delivery.

3. Camera Installation

-Train Arrival Zone: Three key mounting positions along fixed rail routes to fully capture clear images of containers loaded on railcars.

-Reach Stacker Operation Zone: Cameras installed within the working area with adjustable angles to record complete container lifting procedures.

-Yard Exit for Semi-trailers: Cameras mounted at gate exits facing outbound lanes to shoot front or side faces of containers distinctly.

① Train inbound scenario: Chushi applies feature-matching tracking algorithms (e.g. KLT tracker) or deep learning algorithms (e.g. Siamese network) for continuous tracking of containers on rail wagons.

② Reach stacker & semi-trailer scenarios: Given drastic target movement, background subtraction or frame difference algorithms are used for container tracking.

Deep learning models including YOLOv5 and Faster R-CNN perform precise container detection and position localization on video frames.

Detected container images go through preprocessing steps such as cropping and binarization. Chushi’s proprietary OCR technology (referencing Tesseract OCR and Google Cloud Vision API) decodes container IDs. Custom pre-training on container-specific fonts and typography further lifts recognition precision.

1. Four-Tier Software Architecture

-Data Acquisition Layer: Pull real-time video streams from cameras.

-Data Processing Layer: Executes video decoding, preprocessing, object tracking, detection and character recognition.

-Data Storage Layer: Archives processed data and recognition results for later inquiry and statistical analysis.

-Presentation Layer: Human-machine interface displaying live video, identification outcomes and statistical reports.

2. Data Processing Flow

Live video streams are processed frame by frame; key frames are extracted for tracking, detection and OCR. All identified data is written into databases for subsequent data analysis. Standard API interfaces are reserved for third-party system docking and integration.

1.Occlusion: Multi-angle camera layout plus algorithm-based occlusion estimation and data compensation.

2.Irregular Illumination: Image enhancement approaches including histogram equalization and contrast stretching to optimize picture quality.

3.Real-time Performance: Algorithm optimization to enable instant video processing and result output.