基于异构算力平台的视频图像算力算法调度研究

FORMAT DESCRIPTION OF COMPUTER APPLICATIONS AND SOFTWARE

摘要:

针对公安视频监控系统中长期存在的“算力孤岛”与“应用烟囱”问题，本研究提出并实现了基于分层解耦架构的智能算力调度平台。该平台的核心在于构建自主可控的国产化AI算力底座，基于Ascend系列AI处理器，提供模块、板卡、小站、服务器和集群等丰富形态，支持异构计算的端-边-云全场景部署，实现高效的视频图像实时分析和资源池化。并创新性地融合Deep Q-Network (DQN) 强化学习算法与二级协同调度机制，以实现异构算力资源的动态适配与高效利用。通过精细化的算法仓管理、任务智能编排以及资源池化技术，平台能够有效整合NVIDIA GPU、华为NPU等多元异构算力。实验结果表明，该平台可稳定支撑500-1000路视频流的实时解析，任务响应时间显著降低40%，算力利用率提升35%以上。本研究不仅填补了异构算力调度在公安实战应用中的理论空白，还通过引入DQN优化动态决策，为强化学习在边缘计算环境中的应用提供了新范式，具有重要的理论创新与实践价值，尤其在推动视频解析从“人工看”向“智能看”的转型中展现出显著潜力。

Abstract:

In response to the long-standing problems of "computing islands" and "application silos" in the public security video surveillance system, this study proposes and implements an intelligent computing power scheduling platform based on a hierarchical decoupling architecture. The core of this platform lies in building an independently controllable domestic AI computing power foundation, using the Huawei Atlas AI computing platform as the technical base, based on the Ascend series of AI processors, providing various forms such as modules, boards, small stations, servers, and clusters, supporting the end-to-edge-to-cloud full-scenario deployment of heterogeneous computing, and achieving efficient real-time video image analysis and resource pooling. Additionally, it innovatively integrates the Deep Q-Network (DQN) reinforcement learning algorithm and the secondary collaborative scheduling mechanism to achieve dynamic adaptation and efficient utilization of heterogeneous computing resources. Through refined algorithm warehouse management, intelligent task scheduling, and resource pooling technology, the platform can effectively integrate diverse heterogeneous computing resources such as NVIDIA GPUs and Huawei NPU. Experimental results show that this platform can stably support the real-time analysis of 500-1000 video streams, with a significant 40% reduction in task response time and an increase of over 35% in computing power utilization. This study not only fills the theoretical gap in heterogeneous computing power scheduling for public security practical applications but also introduces DQN to optimize dynamic decision-making, providing a new paradigm for the application of reinforcement learning in edge computing environments. It has significant theoretical innovation and practical value, and particularly shows great potential in promoting the transformation of video analysis from "manual viewing" to "intelligent viewing".