【作者】 陈畅；

【导师】 李晓磊；

【作者基本信息】 山东大学 ， 控制工程（专业学位）， 2020， 硕士

【摘要】 近年来,随着计算机视觉的需求程度逐渐提升,目标检测与跟踪成为重要的研究方向之一,但是仍存在一些发展空间。现有的主流目标检测网络为了达到高精确度的要求,往往有着复杂的网络结构,这导致了网络模型参数过大、训练时间过长和对计算机性能要求过高等问题,使得很多前沿的研究成果难以应用于移动设备。面对以上问题,提出了一种基于SSD网络的改进版Deep-SORT算法的轻量化目标检测与跟踪网络FMDeep-SORT,力求在速度、精度和模型大小之间做到有效地权衡。首先,借鉴了 MobileNetV2/V3和ShuffleNetV2的轻量化网络结构,设计出参数量为3.86M的轻量化网络MS-Net,之后结合倒残差块构建轻量化目标检测网络New-SSDLite,经实验证明,本文设计的轻量化目标检测网络PASCAL VOC 2007数据集上,平均准确率达到77.56%,取得了较好的效果;然后,将目标跟踪网络Deep-SORT中原有的目标检测模块Faster-RCNN网络优化为轻量化目标检测网络New-SSDLite,并与均值漂移算法和卡尔曼滤波算法相结合,优化了目标检测模块和跟踪模块,构建轻量化目标跟踪网络FMDeep-SORT,经实验证明,该轻量化目标跟踪网络在目标发生遮挡时,与目前主流的目标跟踪算法比较,仍能取得较好的结果;最后,将该目标检测与跟踪网络应用于超市监控中,对统计调查超市的回购人数具有很大的参考价值。更多还原

【Abstract】 In recent years,with the increasing demand for computer vision,target detection and tracking become one of the important research directions,but there is still some room for development.In order to meet the requirements of high accuracy,there are complex network structures in existing target detection networks.And this leads to problems such as large network model parameters,long training time and high requirements on computer performance,which make many research results unable to be applied to mobile devices.Facing the above problems,this paper processes a lightweight networks for target detection and tracking based on Deep-SORT and SSD networks,named FMDeep-SORT,which can achieve an effective balance between speed,accuracy and size of the model.Firstly,referring to the network structures of MobileNetV2/V3 and ShuffleNetV2,one lightweight network MS-Net is designed.And the parameter amount is only 3.86M.And then combining with inverted residual networks to construct a lightweight target detection network New-SSDLite.Experiments prove that the accuracy rate is 77.56%on the PASCAL VOC 2007 dataset.Secondly,the original target detection module(Faster-RCNN)in Deep-SORT is replaced with New-SSDLite,and combined with the Mean Shift algorithm and Kalman Filtering algorithm,and then a lightweight target tracking network FMDeep-SORT is constructed.Experiments show that FMDeep-SORT achieves better results than the current target tracking algorithms when the target is occluded.Finally,the applicaton of FMDeep-SORT networks in supermarket monitoring has great reference value for counting the number of supermarket repurchases.更多还原