基于SSD网络模型的多目标检测算法  

Multiple target detection algorithm based on SSD networks

作　　者：蔡汉明[1] 赵振兴[1] 韩露[1] 曾祥永 

机构地区：[1]青岛科技大学机电工程学院,山东青岛266042 [2]北京盛开互动科技有限公司,北京100089

出　　处：《机电工程》2017年第6期685-688,共4页Journal of Mechanical & Electrical Engineering

摘　　要：针对现代化工厂中视觉机器人或智能终端处理多目标检测算法的计算任务繁重、运算速度较慢等问题,将网络通信技术应用到算法处理中进行了在线检测。对TCP/IP协议进行了研究,建立了智能终端和云端之间的关系,提出了将智能终端采集到的图像数据进行预处理然后使用基于TCP的Socket多线程通信方式将图像数据送入云端,在云端的多台计算机上同时使用SSD网络模型的多目标检测算法进行了并行处理,并将结果传回智能终端。利用计算机单机与智能终端在线检测在处理时间上进行了对比试验。试验结果表明:在线检测速度稍慢,但已满足实际需求;智能终端在线检测降低了对智能机器人终端硬件的要求,回收的数据可以再利用,并且可以实现算法动态升级。Aiming at the problems of heavy computational tasks and slow speed of operation during visual robots or intelligent terminal han-dling multiple target detection algorithm in modern factory, network communications technology was investigated. In order to establish the relationship between the intelligent terminal and the cloud, TCP/IP protocol were researched, and this method was proposed that the image data collected from the terminals should be preprocessed and sent to the cloud by Socket multi-thread communication based on the TCP, then the image data should be processed in the cloud with multi-target detection of SSD network model, and the results were returned to the termi-nal at last. This algorithm was tested on the processing time by computer LAN or intelligent terminal online on the processing time. The results indicate that on-line measuring is slower, but it meets the practical needs. Intelligent terminal online detection reduces the require-ments for the robot terminal hardware, and the data recycled can be reused, besides, the method can also make the algorithm dynamically be upgraded.

关 键 词：目标检测 卷积神经网络 SSD 智能机器人 SOCKET网络通信 

分 类 号：TP242.6[自动化与计算机技术—检测技术与自动化装置]