摘要
为了解决传统运动车辆检测方法准确性不高、实时性不足的问题,提出基于时空融合加速的Fast RCNN运动车辆快速检测算法。不同于传统方法仅考虑运动车辆的单一特性,所提方法综合考虑运动车辆目标的时域运动特性和空域相关特性。首先基于时域运动特性设计自适应动态背景估计方法,实时估计背景图像,之后基于空域特征利用形态学滤波快速提取疑似目标区域,最后在vgg16模型基础上设计Fast RCNN深度学习网络,对疑似目标区域进行精准真假判别,避免了全图冗余判别。仿真证明,相对于传统方法,该算法能够有效提高检测准确性,且时效性较高,可应用于实时性要求较高的场景。
In order to improve real timing and precision of traditional detection methods for moving vehicle,a fast detection method for moving vehicle target based on spatial-temporal fusion and Fast RCNN is proposed in this paper.Unlike many classical methods adopting only a single feature,motion feature in time domain and spatial correlation feature are used in the proposed method.Firstly,novel adaptive dynamic background estimation is designed,and the background image can be calculated in real time.After that,the moving areas are got by morphological filter based on spatial feature,which are regarded as suspected targets.Finally,Fast RCNN deep-learning network based on vgg16 model is constructed.The fast RCNN network is used to effectively reject false targets from suspected targets rather than full image,which avoids the redundant calculations through the full image.The simulation results indicate that the proposed method can improve real timing and precison of vehicle detection in comparison with the traditional approaches.It proves that the proposed method is more useful for those scenes with high real-time requirements.
作者
陈玉敏
李淼
房晓丽
Chen Yumin;Li Miao;Fang Xiaoli(School of Electronic Information,Hunan Institute of Information Technology,Changsha 410000,China;School of Electronic Science,National University of Defense Technology,Changsha 410000,China)
出处
《电子测量技术》
2020年第3期139-145,共7页
Electronic Measurement Technology
基金
航天系统部专用技术预研基金(30503050503)
2018年湖南省教育厅科学研究项目(18C1583)
湖南信息学院2018年度校级科研项目(XXY018YB02)资助。
