基于Jetson TX2的轻量化SSD模型设计  

Lightweight SSD Model Design Based on Jetson TX2

作　　者：黄建行 谢恒 周益天 张新良[1,2] 郑亚昕 HUANG Jianxing;XIE Heng;ZHOU Yitian;ZHNAG Xinliang;ZHENG Yaxin(School of Electrical Engineering and Automation,Henan Polytechnic University,Jiaozuo 454000,China;Henan International Joint Laboratory of Intelligent Equipment Direct Drive Technology and Control,Jiaozuo 454003,China;Zhengzhou Branch,Agricultural Bank of China,Zhengzhou 450000,China;Zhoushan Yangwangna New Technology Co.,Ltd.,Zhoushan 316041,China)

机构地区：[1]河南理工大学电气工程与自动化学院,河南焦作454000 [2]河南省智能装备直驱技术与控制国际联合试验室,河南焦作454003 [3]中国农业银行郑州分行,河南郑州450000 [4]舟山洋旺纳新科技有限公司,浙江舟山316041

出　　处：《自动化仪表》2025年第4期80-85,共6页Process Automation Instrumentation

基　　金：国家自然科学基金资助项目(U1404612);河南省高校基本科研业务费专项基金资助项目(NSFRF210305);河南省科技攻关基金资助项目(222102210274);大学生创新创业训练计划基金资助项目(202310460058)。

摘　　要：单触发多盒检测器(SSD)模型为图像目标检测提供了高效的解决方案,但增加网络深度以提高精度的方式增加了便携嵌入式设备的负荷,在实际应用场景中受到限制。考虑目标检测实时性的需求,提出了一种轻量化SSD模型。利用金字塔池化模块和多尺度信息融合,替代SSD网络的特征提取模块。金字塔池化模块利用多组池化核在不同尺度上提取特征,并通过横向连接的方式将多个尺度信息进行融合,在保证检测精度的同时,降低嵌入式设备的负荷。给出了在嵌入式设备Jetson TX2上的实现方案。与原SSD模型相比,轻量化SSD模型在嵌入式设备Jetson TX2上的帧率可达13帧/s,并且能够以96%的准确率正确识别佩戴口罩的人员,验证了模型的有效性。该模型为无人驾驶、无人机检测等领域提供了一种实现方案。The single shot-multibox detector(SSD)model provides an efficient solution for image target detection,but increasing the network depth to improve the accuracy increases the load on portable embedded devices and is limited in practical application scenarios.Considering the demand for real-time target detection,a lightweight SSD model is proposed.A pyramid pooling module and multi-scale information fusion are utilized to replace the feature extraction module of the SSD network.The pyramid pooling module uses multiple pooling kernels to extract features at different scales and fuses the multi-scale information by lateral connection to reduce the load on the embedded device while ensuring the detection accuracy.An implementation strategy on the embedded device Jetson TX2 is given.Compared with the original SSD model,the lightweight SSD model achieves a frame rate of up to 13 frame/s on the embedded device Jetson TX2 and can correctly recognize a person wearing a mask with 96%accuracy,which validates the effectiveness of the model.The model provides an implementation strategy for areas such as unmanned and drone detection,etc.

关 键 词：深度学习 单触发多盒检测器 便携嵌入式设备 实时性 轻量化 金字塔池化 多尺度信息融合 Jetson TX2 

分 类 号：TH-39[机械工程]