基于改进YOLOv3的机载平台目标检测算法  

Improved YOLOv 3 Based Target Detection Algorithm for Airborne Platform

作　　者：严开忠 马国梁[1] 许立松 尚海鹏 于睿 YAN Kaizhong;MA Guoliang;XU Lisong;SHANG Haipeng;YU Rui(Nanjing University of Science and Technology,Nanjing 210094 China)

机构地区：[1]南京理工大学,南京210094

出　　处：《电光与控制》2021年第5期70-74,共5页Electronics Optics & Control

基　　金：国家自然科学基金青年基金(11302106)。

摘　　要：针对小型智能侦察无人机机载平台存在的计算力受限、检测速度较慢的问题,提出了一种基于YOLOv3改进的目标检测算法。首先引入深度可分离卷积改进YOLOv3的骨干网络,降低网络的参数和计算量,提高算法的检测速度,再根据机载视角下目标形状的特点,预置K-means产生先验框的初始聚类中心,并在边框回归中引入CIoU损失函数,将DIoU与NMS结合,改善YOLOv3对密集目标的漏检问题,最后再通过TensorRT优化加速后部署到英伟达Jetson TX2机载计算平台。实验结果表明,所改进的算法在验证集上的平均精度均值(MAP)达到了82%,检测速度从3.4帧/s提升到16帧/s,满足实时性要求。To overcome the problems of limited calculation power and slow detection speed of the small intelligent reconnaissance UAV platforms an improved target detection algorithm based on YOLOv3 is proposed.First of all depthwise separable convolution is introduced to improve the backbone network of YOLOv3 which greatly reduces the quantity of parameters and calculation cost of the network and improves the detection speed of the algorithm.Then according to the characteristics of the target shape under the perspective of the airborne platform the initial clustering center of K-means is preset when generating prior box and CIoU loss function is introduced in the box regression.DIoU is combined with NMS to reduce the missed detections for dense targets.Finally the improved model is optimized and speeded up by TensorRT and deployed to the NVIDIA Jetson TX2 airborne computing platform.The experimental results show that the Mean Average Precision(MAP)of the improved algorithm on the verification set reaches 82%and the detection speed is increased from 3.4 to 16 frames which can meet the real-time requirements.

关 键 词：目标检测 侦察无人机 YOLOv3 深度可分离卷积 DIoU TensorRT 

分 类 号：V247[航空宇航科学与技术—飞行器设计]