摘要
随着无人机技术和深度学习技术的发展,基于深度学习的多目标检测算法在工业无人机中得到了广泛的应用。针对目前基于深度学习的多目标检测算法占用大量计算量资源,难以在算力有限的中小型无人机平台上实时运行的问题,分析了深度学习算法在低功耗CPU上的耗时,提出一种卷积神经网络计算优化方法。在机载计算机中进行仿真,结果表明在检测效果基本不变的条件下,算法帧率达到了56FPS,实现了无人机平台上的实时多目标检测。
The development of quadrotor drone and deep learning, target detection algorithms based on deep learning have been widely used in the drone industry. Existing deep learning-based detection algorithms cannot be run in real time on drone onboard computer with limited computing resources. This paper analyzesd the distribution of run time of deep learning algorithm on low power CPU, and proposed a computational optimization method for convolutional neural networks. The experimental results in drone onboard computer show that the detection speed can reach 56 FPS while the detection result keeps unchanged.
作者
朱壬泰
胡士强
ZHU Ren-tai;HU Shi-qiang(School of Aeronautics and Astronautics of Shanghai Jiaotong University,Shanghai 200240,China)
出处
《计算机仿真》
北大核心
2020年第4期57-61,共5页
Computer Simulation
关键词
深度学习
多目标检测
四旋翼的
Deep learning
Multi-target detection
Quadrotor
作者简介
朱壬泰(1994-),男(汉族),安徽桐城人,硕士研究生,主要研究领域为无人机视觉导航;胡士强(1969-),男(汉族),河北石家庄人,教授,博士研究生导师,主要研究领域为信息融合技术与图像理解与分析。
