Abstract
As global urbanization accelerates, issues like road congestion and traffic accidents become increasingly severe. Intelligent Transportation Systems (ITS) leverage advanced information and communication technologies to create efficient, safe, and eco-friendly transportation frameworks. Image processing is a core technology in ITS, providing essential data for perception, detection, and path planning. However, the massive data generated by modern ITS applications—such as autonomous driving—strains centralized cloud architectures. Edge computing addresses this by offloading computation to the network edge, meeting the ultra-low latency requirements of real-time services. This paper reviews the state-of-the-art in image processing and Multi-access Edge Computing (MEC) for ITS, highlighting research hotspots and future trends.
Introduction: The Data Challenge in Modern ITS
Urban growth has brought traditional transportation systems to a breaking point, characterized by rising congestion and safety concerns. ITS aims to mitigate these issues through advanced sensing. However, modern applications like autonomous driving generate an estimated 300 TB of data per vehicle annually. Centralized cloud computing is often insufficient for safety-critical tasks due to high latency. This review explores how sophisticated computer vision algorithms can be efficiently deployed using a distributed "Cloud-Edge-Terminal" architecture to ensure real-time, reliable perception.
Core Technologies: Image Processing in ITS
Image processing techniques are fundamental to transforming raw visual data into actionable intelligence. The paper categorizes these into several key areas:
Enhancement and Restoration
To maintain reliability in adverse conditions, techniques are employed to handle poor lighting, fog, and rain. The review contrasts traditional hand-crafted methods, such as those based on Retinex theory, with modern deep learning approaches like Generative Adversarial Networks (GANs) and CNN-based weak-light enhancement.
Object Detection and Segmentation
The field is currently split between "two-stage" detectors (e.g., Faster R-CNN) that prioritize high accuracy and "one-stage" detectors (e.g., YOLO, SSD) optimized for real-time speed. Furthermore, pixel-level classification through semantic segmentation (using models like FCN, SegNet, or U-Net) is essential for identifying lane lines and complex road layouts.
The Edge Computing Paradigm
Multi-access Edge Computing (MEC) acts as a localized "brain" for the roadway, significantly reducing the distance data must travel.
- Task Offloading: Strategies determine whether a computation should occur on the vehicle or at a Roadside Unit (RSU) to optimize for latency and energy consumption.
- Data Caching: Frequently accessed resources, such as high-definition maps or local traffic alerts, are stored at the edge to minimize backhaul traffic and improve response times.
Key Insights: The Shift to CVIS
A major takeaway is the transition from single-vehicle intelligence to Cooperative Vehicle-Infrastructure Systems (CVIS). Single-vehicle systems often face bottlenecks due to sensor costs and the "black box" nature of AI. CVIS alleviates this by sharing the perception burden with intelligent roadside infrastructure, allowing vehicles to effectively "see" beyond their immediate line of sight.
Performance and Results
The paper highlights significant quantitative milestones:
- Detection Efficiency: While two-stage detectors offer high precision (e.g., 78.8% mAP on VOC07), one-stage models like YOLOv4 have become industry standards for real-time deployment, achieving 43.5% mAP on MS COCO at 65 FPS.
- Latency Reduction: Implementing edge computing can reduce communication latency by approximately 80% compared to traditional cloud-only architectures.
- Reliability: License Plate Recognition (LPR) systems have reached over 98% accuracy, even in challenging nighttime environments.
Future Outlook and Challenges
Despite these advancements, several hurdles remain for widespread deployment:
- Environmental Robustness: Models still struggle with extreme weather like heavy snow or unpredictable "long-tail" traffic scenarios.
- Resource Optimization: Balancing limited computational power and energy at edge nodes during peak traffic remains a complex problem.
- Security: Data sharing between vehicles and infrastructure introduces risks related to data spoofing and privacy.
Future research is expected to focus on cross-modal data fusion—integrating LiDAR, Radar, and Camera data at the edge—and developing more energy-efficient, "green" edge computing algorithms.
Summary updated