Abstract
Vehicle re-identification (Re-ID) is a cornerstone of autonomous driving perception, yet traditional methods are often ill-suited for the 360° surround-view systems common in modern platforms. These systems utilize multiple fisheye cameras, introducing significant lens distortion and extreme viewpoint variations. This paper proposes a specialized Re-ID pipeline that addresses these challenges through a novel Quality Evaluation Mechanism for single-camera tracking and a geometry-aware spatial constraint strategy for multi-camera association. Evaluated on a new large-scale fisheye dataset, the proposed system achieves state-of-the-art identity consistency while maintaining real-time performance (30-42 FPS).
Introduction
Autonomous vehicles typically rely on a suite of surround-view cameras to maintain a complete 360° awareness of their environment. While vehicle Re-ID has been extensively studied in fixed-position surveillance contexts, the transition to mobile, multi-fisheye systems introduces unique hurdles. Standard appearance-based matching often fails when a vehicle moves from the center of a fisheye frame to its distorted edge, or when it transitions between cameras with vastly different perspectives.
Problem Statement: The Surround-View Challenge
The primary difficulties in surround-view Re-ID stem from two factors:
- Single-Camera Distortion: Fisheye lenses cause a vehicle's visual signature to change dramatically based on its position in the frame.
- Multi-Camera Variance: The viewpoint shift between adjacent cameras (e.g., Front to Left) is so extreme that traditional Re-ID networks struggle to maintain identity consistency.
Methodology
The proposed pipeline decomposes the Re-ID task into single-camera tracking and multi-camera association.
Single-Camera Tracking with Quality Evaluation
To mitigate the effects of fisheye distortion, the authors enhance the SiamRPN++ tracker with a custom Quality Evaluation Mechanism. This mechanism relies on two key metrics:
- Center Drift ($IoU_R$): Recognizing that bounding box shapes are unstable in fisheye views, $IoU_R$ prioritizes the stability of the target's center point over simple area overlap.
- Occlusion Coefficient ($OC$): This metric quantifies target visibility, allowing the system to maintain tracklets during temporary occlusions or truncations.
Multi-Camera Association and Spatial Constraints
For cross-camera matching, the system utilizes an attention-enhanced Re-ID network (BDBnet) combined with a "Spatial Constraint Strategy." By leveraging the fixed geometry of the camera rig, the authors estimate the expected position of a vehicle in a neighboring camera's view using wheel grounding points. This geometric projection creates "uncertainty ellipses" that significantly narrow the search space, filtering out false positives before appearance-based matching occurs.
Experimental Results
Due to the lack of suitable public datasets, the authors introduced a new large-scale annotated fisheye dataset comprising 22,190 images across 36 sequences.
Key findings include:
- Tracking Stability: The introduction of $IoU_R$ and Re-ID confidence metrics improved Identity Consistency (IC) in the front camera from 0.82 to 0.94.
- Association Accuracy: The combination of the Attention Module and Spatial Constraints boosted multi-camera matching accuracy to a range of 0.94–0.97.
- Real-time Performance: The system maintains a processing speed of 30-42 FPS, making it viable for deployment on autonomous driving hardware.
Conclusion
This work demonstrates that robust vehicle Re-ID in surround-view systems requires moving beyond pure appearance-based methods. By integrating geometric constraints and distortion-aware tracking metrics, the proposed pipeline successfully maintains vehicle identities across complex 360° environments. Future work will explore enhancing robustness on uneven terrain where geometric projections may be less precise.
Summary updated