Vehicle Violations Detection Using Deep Learning
No Thumbnail Available
Date
2025
Authors
Journal Title
Journal ISSN
Volume Title
Publisher
Saudi Digital Library
Abstract
Overweight and over-dimension commercial vehicles accelerate pavement deterioration, increase crash severity, and impose significant operational burden on enforcement agencies. However, conventional enforcement approaches—static weighbridges and manual roadside inspections—are disruptive to traffic flow, labor-intensive, and difficult to scale for continuous monitoring on high-volume highways. This dissertation presents a Hybrid Vision/Virtual Weigh-in-Motion (V-WIM) framework that integrates certified Weigh-in-Motion (WIM) measurements with computer vision and deep learning to support automated, enforcement-oriented truck violation detection at highway speeds. The proposed pipeline comprises three key components. First, a YOLO/CNN perception stack detects commercial vehicles and axles, producing per-frame observations that include vehicle localization and measurement cues. Second, a Gated Recurrent Unit (GRU) temporal smoothing module stabilizes noisy frame-level dimension estimates to yield consistent per-vehicle length/width/height outputs suitable for compliance assessment. Third, a Graph Neural Network (GNN) fusion and association layer constructs a consistency graph over candidate WIM records and vision tracklets and performs graph-based inference to associate cross-modal observations and produce a final violation score and decision under legal thresholds. A standardized data orchestration layer outputs synchronized event records in structured formats (e.g., CSV/YAML), enabling repeatable training, evaluation, and auditing of results. The framework is evaluated across detection, association, and decision stages using precision, recall, F1-score, AP, ROC-AUC, confusion matrices, and threshold sweeps. Results indicate that combining temporal smoothing with graph-based fusion improves measurement stability and decision reliability compared with a YOLO-only baseline, supporting scalable, real-time monitoring and more defensible violation decisions for intelligent transportation enforcement.
Vehicle Violations Detection.
Description
Overweight and over-dimension trucks contribute significantly to pavement damage, higher maintenance costs, and elevated safety risk on highway networks. Yet many current enforcement practices still depend on static weigh stations or manual roadside checks, which interrupt traffic, require substantial staffing, and provide limited coverage for continuous monitoring. This dissertation develops an automated, enforcement-focused solution using deep learning and multi-sensor integration, referred to as a Hybrid Vision/Virtual Weigh-in-Motion (V-WIM) framework, to identify truck violations at normal travel speeds. The proposed system combines sensing and intelligence across three stages. First, a real-time computer vision module based on YOLO and CNN components detects trucks and axles from roadside cameras and extracts frame-level measurement cues. Second, a temporal modeling stage uses a Gated Recurrent Unit (GRU) to reduce noise and instability in the visual measurements, converting short-term frame estimates into a single, robust per-vehicle dimension profile. Third, a Graph Neural Network (GNN) association and fusion layer links vision tracklets with corresponding WIM records (and optional 3D sizing inputs such as LiDAR-derived dimensions) by learning relationships constrained by timing and vehicle motion consistency. This graph-based reasoning produces a final violation likelihood score and a decision aligned with regulatory thresholds. To ensure repeatability and auditability, the workflow is supported by a standardized data pipeline that synchronizes sensor streams and exports structured event files for training and evaluation. Performance is assessed across detection, matching, and decision tasks using precision, recall, F1-score, average precision, ROC-AUC, confusion matrices, and threshold analyses. The results demonstrate that temporal smoothing and graph-based fusion improve measurement stability and strengthen final enforcement decisions compared with a vision-only baseline, enabling scalable and defensible automated monitoring of truck violations.
Keywords
Vehicle Violations Detection, Overweight Vehicle Enforcement, Oversize / Over-Dimension Vehicle Detection, Deep Learning–Based Enforcement, Vision-Based Weigh-in-Motion (V-WIM), Weigh-in-Motion (WIM) Systems, Computer Vision for Intelligent Transportation Systems (ITS), Truck and Axle Detection, YOLO (Real-Time Object Detection), Temporal Smoothing (GRU), Multi-Sensor / Cross-Modal Fusion, Graph Neural Networks (GNN) for Data Association, Vehicle Dimension Estimation (3D Sizing), Automated Violation Decision / Rule-Based Thresholding, Real-Time Highway Monitoring, Sensor Synchronization and Event Matching, Enforcement-Grade Confidence Scoring, Intelligent Traffic Monitoring / Smart Mobility