基于轻量级级联网络的边缘端人员身份识别系统

Edge-End Personnel Identification System Based on Lightweight Cascaded Networks

摘要: 本文提出了一种基于任务级级联融合的人员身份识别的方法，解决传统并行架构中计算冗余、延时高、不稳定的问题。采用了“YOLOv4人体检测，ROI人脸区域提取与YuNet人脸检测，InsightFace身份识别，状态管理”四个步骤作为整体的顺序，并且用前序模块输出的结果来限制后序处理的范围，实现了任务信息在任务间的有序传递以及任务资源的最佳分配。适用于对于实时性和准确率均有一定要求的安防监控、智能考勤等场景。该方法通过融合三种技术手段来提高性能，一是引入级联约束，在人体检测框内搜寻人脸，把待处理像素减少了约62%，推理速度加快了近2.6倍；二是使用了多帧识别融合方案，基于10秒滑动窗口内的历史识别队列，结合身份频次和置信度的滤波规则去除了由于光照、姿态变化带来的错误识别，整体准确率提升约12%；三是通过“已识别，追踪中，已确认”的三级状态实现结果的连贯性以及业务适配性。经过实验得知，在Intel 16核CPU平台下，该系统级联融合整体处理速度达到17.06 FPS (处理延时58.6 ms)，端到端识别准确率达到93.55%，在WIDER FACE数据集上人脸检测精度达85.45%，在MOT17数据集上追踪MOTA达0.32，无需使用GPU就可以完成实时监控任务，具有较高的实用价值以及可部署于边缘端的特点。

Abstract: This paper proposes a personnel identification method based on task-level cascaded fusion, solving the problems of computational redundancy, high latency, and instability in traditional parallel architectures. It adopts a four-step sequence: “YOLOv4 human detection, ROI face region extraction and YuNet face detection, InsightFace identification, and state management,” and uses the output of the preceding modules to limit the scope of subsequent processing, achieving orderly transfer of task information between tasks and optimal allocation of task resources. This method is suitable for scenarios such as security monitoring and intelligent attendance systems that require both real-time performance and accuracy. This method improves performance by integrating three techniques: First, it introduces cascaded constraints to search for faces within the human detection bounding box, reducing the number of pixels to be processed by approximately 62% and accelerating inference speed by nearly 2.6 times. Second, it uses a multi-frame recognition fusion scheme, based on a historical recognition queue within a 10-second sliding window, combining filtering rules based on identity frequency and confidence to remove erroneous recognitions caused by changes in lighting and pose, improving overall accuracy by approximately 12%. Third, it achieves consistency and business adaptability through a three-level status system: “Recognized,” “Tracking,” and “Confirmed.” Experiments show that on an Intel 16-core CPU platform, the system achieves an overall processing speed of 17.06 FPS (processing latency of 58.6 ms), an end-to-end recognition accuracy of 93.55%, a face detection accuracy of 85.45% on the WIDER FACE dataset, and a MOTA tracking accuracy of 0.32 on the MOT17 dataset. It can complete real-time monitoring tasks without using a GPU, demonstrating high practical value and the ability to be deployed at edge computing.