一种基于前视声呐的目标检测与跟踪方法  

A target detection and tracking method based on forward looking sonar

作　　者：刘昊搏 刘铁军[1,2,3] 汪海林[1,2,3,4] 姜志斌 祁胜 LIU Hao-bo;LIU Tie-jun;WANG Hai-lin;JIANG Zhi-bin;QI Sheng(Shenyang Institute of Automation,Chinese Academy of Sciences,Shenyang 110016,China;Institutes for Robotics and Intelligent Manufacturing Chinese Academy of Sciences,Shenyang 110169,China;Key Laboratory of Marine Robotics,Liaoning Province,Shenyang 110169,China;University of Chinese Academy of Sciences,Beijing 100049,China)

机构地区：[1]中国科学院沈阳自动化研究所,辽宁沈阳110016 [2]中国科学院机器人与智能制造创新研究院,辽宁沈阳110169 [3]辽宁省水下机器人重点实验室,辽宁沈阳110169 [4]中国科学院大学,北京100049

出　　处：《舰船科学技术》2021年第11期143-148,共6页Ship Science and Technology

基　　金：国家重点研发计划(2017YFC0305703)。

摘　　要：随着海洋勘探技术的发展,水声图像在海洋开发中占据越来越重要的地位。针对前视声呐在障碍物识别中存在的高噪声以及目标成像不连续等问题,借鉴神经网络中卷积核的思想,将形态学算子视为特殊的卷积核。同时,将帧间差分图像与当前帧所获取图像的特征相融合,提出一种基于特征融合的目标检测算法。而后利用Mean-shift(均值漂移)算法对检测后得到的清晰的障碍物目标进行跟踪。实验结果表明,该算法能够有效地抑制动态背景中的噪声,同时使障碍物目标能够以连续、稳定变化的运动形态,清晰地呈现在连续多帧图像中,并对目标进行有效地跟踪。以"探索1000"AUV在千岛湖采集的声呐视频序列作为实验数据,表明该算法具有广阔的应用前景。With the development of ocean exploration technology,underwater acoustic images occupy an increasingly important position in ocean development.Aiming at the problems of high noise and discontinuous target imaging in the obstacle recognition of forward-looking sonar,the idea of convolution kernel in neural network is used for reference,and the morphological operator is regarded as a special convolution kernel,At the same time,the inter-frame difference image is fused with the features of the image acquired in the current frame,and a target detection algorithm based on feature fusion is proposed.Then,the Mean-shift algorithm is used to track the clear obstacle targets obtained after detection.The experimental results show that the algorithm can effectively suppress the noise in the dynamic background,at the same time,enable the obstacle target to be clearly presented in continuous multi-frame images in a continuous and stable changing motion form,and to effectively track the target.The sonar video sequence collected by "Exploration 1 000" AUV in Qiandao Lake is used as experimental data,which shows that this algorithm has broad application prospects.

关 键 词：前视声呐 均值漂移 形态学算子 障碍物检测 帧差法 

分 类 号：TP391[自动化与计算机技术—计算机应用技术] TN911.73[自动化与计算机技术—计算机科学与技术]