目前矿山AI视频系统主要采用地面服务器进行分析处理，存在视频分析整体响应时延较高、多系统联动延迟、网络带宽资源占用大等问题。针对该问题，设计了一种轻量化、可边缘部署的矿用AI视频边缘计算系统。提出了基于注册机的轻量化软件开发工具包（SDK）框架，对算子进行解耦，以提升算法并行运算能力，降低SDK对存储空间的需求；对YOLOv7的卷积运算进行分组设计，利用恒等映射对Focus主干网络进行优化，以减少运算量并轻量化网络结构，同时引入Transformer中的注意力机制提高检测性能。集成国产化智能芯片和5G通信模块，研制了矿用AI视频服务器，实现了矿山边缘节点部署与计算。实验结果表明：矿用AI视频边缘计算系统响应出色，部署注册机SDK与改进YOLOv7模型后，平均推理时延为28 ms，较React Native+YOLOv7和MobileNet分别减少52%和44%；在各种负载情况下，矿用AI视频服务器的响应时间远低于矿山操作的最低要求。现场工业性测试结果表明：矿用AI视频服务器接入8路摄像仪时，响应时延为51 ms，带宽维持在45 Mbit/s，比使用地面服务器时的时延降低了59%，带宽提高了15%,实现了对井下视频数据的实时、就地分析处理，有效降低了数据传输时延，提高了视频分析的响应速率和处理效率。

Currently, mining AI video systems mainly rely on ground servers for analysis and processing, which leads to issues such as high overall response latency, multi-system linkage delays, and high network bandwidth utilization. To address these issues, a lightweight and edge-deployable mining AI video edge computing system was designed. A lightweight software development kit (SDK) framework based on the register machine was proposed to decouple operators, improving the parallel computing ability of the algorithms and reducing the storage requirements of the SDK. Grouping design of YOLOv7 convolution operation was conducted, and the Focus backbone network was optimized using identity mapping to reduce computation amount and streamline the network structure. Additionally, the attention mechanism from Transformer was incorporated to improve detection performance. The mining AI video server was developed by integrating the domestic intelligent chips and 5G communication modules, enabling the deployment and calculation of mine edge nodes. The experimental results showed that the mining AI video edge computing system had excellent response. After deploying the registration machine SDK and optimizing the YOLOv7 model, the average inference latency was 28 ms, 52% and 44% lower than that of React Native+YOLOv7 and MobileNet, respectively. Under various load conditions, the response latency of the mining AI video server was significantly lower than the minimum requirements for mine operation. The field industrial test results showed that when the mining AI video server was connected to an 8-channel camera, the response latency was 51 ms, and the bandwidth was maintained at 45 Mbit/s. Compared to using a ground server, the response latency was reduced by 59%, and the bandwidth increased by 15%. This enabled real-time and on-site analysis and processing of underground video data, effectively reducing the data transmission delay and improving the response rate and processing efficiency of video analysis.