Phased Feature Extraction Network for Vehicle Search Tasks Based on Cross-Camera for Vehicle-Road Collaborative...

Area of Science:

• Computer Vision
• Artificial Intelligence
• Autonomous Systems

Background:

• Vehicle search integrates vehicle detection and re-identification (Re-ID) for real-world image analysis.
• Existing networks are often designed for pedestrian search, necessitating adaptation for vehicle-specific challenges like perspective variations.
• The DAIR-V2X dataset was leveraged to create DAIR-V2XSearch, the first cross-camera dataset for vehicle search.

Purpose of the Study:

• To develop a robust network for cross-camera vehicle search, addressing limitations of existing methods.
• To overcome challenges posed by diverse perspectives and imaging conditions in real-world driving scenarios.
• To establish a benchmark for vehicle search performance using the novel DAIR-V2XSearch dataset.

Main Methods:

• Proposed a phased feature extraction network (PFE-Net) utilizing the anchor-free YOLOX framework as a backbone.
• Introduced a camera grouping module within the Re-ID branch to handle perspective and camera disparities.
• Implemented a cross-level feature fusion module to enhance subtle feature extraction and Re-ID precision.

Main Results:

• The PFE-Net demonstrated superior performance on the DAIR-V2XSearch dataset.
• The anchor-free approach mitigated issues with multiple anchor boxes per vehicle ID.
• The camera grouping and feature fusion modules effectively addressed cross-camera challenges.

Conclusions:

• The PFE-Net provides a significant advancement for cross-camera vehicle search tasks.
• The proposed methods enhance the accuracy and robustness of vehicle Re-ID in complex environments.
• This work lays the foundation for improved perception systems in intelligent driving and autonomous vehicles.