AlignDual : Real-Time Multimodal 3D Detection on Resource-Constraint Edge Device via Inter-Stream Cooperation

Multimodal 3D Object Detection (M3DOD) is critical for applications like city surveillance, industrial defect detection, and autonomous driving, yet state-of-the-art algorithms are too resource-intensive for edge devices. While cloud offloading appears to be a solution, we identify a fundamental data misalignment problem inherent to this approach that degrades detection performance. This failure manifests as two intertwined issues: (1) Semantic Misalignment , where uncoordinated, modality-agnostic compression can discard the cross-modal spatial correlations essential for fusion, and (2) Temporal Misalignment , where heterogeneous pipeline delays lead to substantial synchronization bottlenecks and violate real-time constraints. This paper introduces AlignDual , a novel framework that tackles these challenges by establishing a new paradigm: Cross-Modal Co-Design. Instead of treating sensor streams as independent flows, AlignDual establishes two key cooperative mechanisms. First, a semantically-coordinated compression scheme leverages edge-efficient 2D object semantics to guide point cloud sampling at the source, preserving fusion-critical correlations before transmission. Second, a proactive, prediction-based synchronization framework abandons reactive waiting, instead using motion prediction to compensate for latency jitter and reduce synchronization overhead. These mechanisms are orchestrated by a closed-loop optimizer that dynamically adapts to runtime conditions. We implemented and evaluated AlignDual on a real-world testbed. Results show that our system outperforms state-of-the-art cloud-based approaches, improving detection accuracy (mAP) by up to 18.2% while simultaneously increasing the real-time latency compliance rate (CR) by 22.3%.