Edge-Assisted Video Object Detection With Adaptive Resource Allocation and Model Selection

With the rise of applications such as facial recognition and intelligent monitoring, demand for video object detection has surged. To optimize detection performance, existing works primarily focus on either video transmission or processing independently, often overlooking small object detection, which is particularly difficult for resource-limited terminals. To address these limitations, we propose a solution that leverages edge servers to assist video object detection, enabling real-time and high-precision performance under resource constraints. We introduceAdaptive-EVOD, the first edge-assisted video detection system that optimizes communication and computation simultaneously through adaptive resolution and frame rate adjustment, resource allocation, and model selection strategies, with a particular emphasis on small object detection. Specifically,Adaptive-EVODintroduces three core innovations: (1) proposing a detection strategy for small objects to enhance precision; (2) employing deep reinforcement learning to dynamically determine detection location, resolution, and frame rate based on heterogeneous environments; and (3) applying Bayesian optimization to dynamically allocate GPU resources and select Low-Rank Adaptation (LoRA) ranks for detection models adaptively. Experiments on a real-world testbed show thatAdaptive-EVODreduces end-to-end latency by 5.7%–18.3% and improves detection accuracy by 7.8%–23.4% compared to state-of-the-art methods, which highlights the effectiveness ofAdaptive-EVOD, particularly in complex and resource-constrained edge scenarios.