A Lightweight End-to-End Framework for Real-Time Vehicle-Ejected Debris Detection on Edge Devices

Area of Science:

• Computer Vision
• Artificial Intelligence
• Traffic Engineering

Background:

• Vehicle-ejected debris detection is crucial for intelligent traffic enforcement but remains challenging due to small, transient objects.
• Current methods rely on manual review, lacking efficient, real-time solutions and edge-deployable models.

Purpose of the Study:

• To develop a lightweight, hardware-aware detection framework for real-time vehicle-ejected debris monitoring on edge devices.
• To construct a comprehensive dataset for training and evaluating debris detection models in diverse road scenarios.

Main Methods:

• A novel dataset of 4328 annotated images was created, covering various road conditions and lighting.
• A YOLOv8m-based framework was adapted using MobileNetV4 and a Channel Alignment Module for efficiency and edge compatibility.
• The model was quantized to INT8 and deployed on the RDK X5 platform for real-time performance evaluation.

Main Results:

• The proposed model achieved 93.1% mAP50, significantly reducing parameters (13.1M) and GFLOPs (39.6) compared to YOLOv8m.
• INT8 deployment on the RDK X5 platform reached 112.6 FPS with minimal accuracy loss.
• The framework demonstrated a practical accuracy-efficiency advantage for edge-based debris detection.

Conclusions:

• The developed framework offers a practical, edge-deployable solution for real-time vehicle-ejected debris monitoring in traffic enforcement.
• This work lays the foundation for automated detection systems, though further research is needed for event verification and attribution.