基于深度学习的ViSAR多运动目标检测

Moving Target Shadow Detection Approach for ViSAR Using Deep Learning and Multi-Object Tracking Algorithm

视频合成孔径雷达(ViSAR)在地面动目标检测和感兴趣区域(ROI)的动态监测方面具有巨大的潜力。对地面运动目标的检测与跟踪一直是ViSAR的研究热点。针对现有基于深度学习的ViSAR动目标检测方法存在的依赖预训练模型,模型迁移难等问题,本文提出了一种基于深度学习与多目标跟踪(MOT)算法的ViSAR动目标阴影检测方法。该方法首先设计了一种从零开始深度学习的网络模型,实现动目标阴影的单帧检测。为了提高检测性能的鲁棒性,采用了基于卡尔曼滤波和逐帧数据关联的多目标跟踪算法跟踪动目标。实测数据处理结果表明该方法具有良好的检测性能。

Video synthetic aperture radar(ViSAR)presents great potential for ground moving target detection and dynamic surveillance of regions of interest(ROI).Detection and tracking of ground moving targets has been one of the research hotspots of ViSAR.To address the problems of the existing deep learning-based ViSAR moving target detection methods,such as reliance on pre-trained models and difficulty of model migration,a novel moving target shadow detection approach based on deep learning and multi-object tracking(MOT)algorithm is proposed in this paper.We design a network model with deep learning from scratch to achieve single-frame detection of moving target shadows.To improve the robustness of detection performance,a multi-object tracking algorithm based on the Kalman filter and frame-by-frame data association algorithm is employed to track moving targets.Experiments of measured data are performed and have demonstrated that the proposed approach processes an accepted detection performance.