Distributed model cache-driven device-cloud collaboration in adaptive video analytics

In video analysis, data drift is one of the main issues affecting system inference performance. To address this, this paper proposes a mobile-oriented edge-cloud collaborative continuous learning framework. Specifically, this framework deploys a model caching on resource-constrained mobile devices and designs an edge-cloud collaborative hierarchical adaptation strategy that integrates “reuse, fine-tuning, and retraining”. Constructing a prefix tree storage structure based on domain attributes enables rapid model retrieval, thereby reducing cloud queuing delays during adaptations. Evaluation results demonstrate that this framework outperforms existing state-of-the-art solutions in object detection and image classification tasks, achieving a 34.5-times reduction in reuse response time and a 1.99-times reduction in retraining time while achieving optimal overall accuracy in multi-device system adaptations. This highlights the framework’s significant potential for enhancing response speed and system scalability.