Abstract:

In order to meet the requirements for safe, accurate and efficient processing of security image data, this paper designs a real-time object detection framework in a low-latency and high-reliability edge computing environment, which is divided into terminal network layer, edge server layer and cloud layer. At the terminal network layer, the BING（Binarized Normed Gradients） algorithm is used to distinguish the object area and background area of the image, and selective compression is applied to different areas to reduce the amount of data transmission. At the edge server layer, an improved SSD（Single Shot Multibox Detector）-based algorithm 2FSSD is proposed to address the problem that small objects in security images are easy to miss detection. The deep level features with high semantic information and low level features in SSD backbone network are fused to form a new feature extraction layer for object detection. Experiments show that the 2FSSD algorithm has a detection accuracy of 79.5% and a detection speed of 27.3 frame/s in VOC data set. Therefore, the research approach can meet the real-time requirements of security image processing while ensuring the detection accuracy.

Key words: edge computing, object detection, small object, SSD

摘要：

为了满足对安防图像数据安全准确高效的处理需求，在低延迟、高可靠性的边缘计算环境中设计了一种实时性目标检测框架，按层次分为终端网络层、边缘服务器层与云层。在终端网络层采用BING（Binarized Normed Gradients）算法区分图像的目标区域与背景区域，对不同区域采取选择性压缩以减少数据传输量。在边缘服务器层，针对安防图像中小目标易漏检的问题提出了基于SSD（Single Shot Multibox Detector）的改进算法2FSSD，将SSD主干网络中具有高语义信息的深层特征与低层特征进行融合，形成新的特征提取层进行目标检测。实验表明，2FSSD算法在VOC数据集中检测精度可达79.5%，检测速度为27.3 frame/s。因此该研究方法可以在保证检测准确度的同时满足安防图像处理的实时性要求。

关键词: 边缘计算, 目标检测, 小目标, SSD