Cross-camera vehicle trajectory reconstruction in roadside surveillance networks with partial and minimal overlaps

Abstract Cross-camera vehicle trajectory reconstruction is essential for roadside perception systems supporting traffic analysis, safety evaluation, and infrastructure monitoring. In practical deployments, roadside cameras often exhibit partial or minimal overlaps, providing limited spatiotemporal continuity and making cross-camera association challenging. This paper proposes a global trajectory reconstruction framework designed for such roadside surveillance networks. Single-camera trajectories are first extracted and projected into a unified road-plane coordinate system through calibration, enabling metric-level comparison across views. Appearance features and motion cues are then jointly exploited in a multi-stage association strategy that integrates spatiotemporal feasibility and visual similarity. Matched trajectories are stitched and refined using quality-aware alignment and short-gap compensation to generate continuous and physically plausible cross-camera trajectories. Real-world experiments with roadside cameras, RTK-equipped vehicles, and UAV-based trajectory data demonstrate that the proposed method achieves high association accuracy and reconstruction precision in both single- and multi-vehicle scenarios while maintaining low computational overhead.