Camera-based object tracking systems in a given closed environment lack privacy and confidentiality.In this study,light detection and ranging(LiDAR)was applied to track objects similar to the camera tracking in a clos...Camera-based object tracking systems in a given closed environment lack privacy and confidentiality.In this study,light detection and ranging(LiDAR)was applied to track objects similar to the camera tracking in a closed environment,guaranteeing privacy and confidentiality.The primary objective was to demonstrate the efficacy of the proposed technique through carefully designed experiments conducted using two scenarios.In Scenario I,the study illustrates the capability of the proposed technique to detect the locations of multiple objects positioned on a flat surface,achieved by analyzing LiDAR data collected from several locations within the closed environment.Scenario II demonstrates the effectiveness of the proposed technique in detecting multiple objects using LiDAR data obtained from a single,fixed location.Real-time experiments are conducted with human subjects navigating predefined paths.Three individuals move within an environment,while LiDAR,fixed at the center,dynamically tracks and identifies their locations at multiple instances.Results demonstrate that a single,strategically positioned LiDAR can adeptly detect objects in motion around it.Furthermore,this study provides a comparison of various regression techniques for predicting bounding box coordinates.Gaussian process regression(GPR),combined with particle swarm optimization(PSO)for prediction,achieves the lowest prediction mean square error of all the regression techniques examined at 0.01.Hyperparameter tuning of GPR using PSO significantly minimizes the regression error.Results of the experiment pave the way for its extension to various real-time applications such as crowd management in malls,surveillance systems,and various Internet of Things scenarios.展开更多
在目标检测领域中,基于交并比(intersection over union, IoU)的系列损失函数存在一定的局限性,使得边界框回归的精度和稳定性有待进一步提升。为此提出了一种基于非线性高斯平方距离的边界框回归损失函数。首先综合考虑了边界框中重叠...在目标检测领域中,基于交并比(intersection over union, IoU)的系列损失函数存在一定的局限性,使得边界框回归的精度和稳定性有待进一步提升。为此提出了一种基于非线性高斯平方距离的边界框回归损失函数。首先综合考虑了边界框中重叠性、中心点距离和长宽比3个因素,将边界框建模为高斯分布;然后提出一种高斯平方距离来衡量概率分布之间的差距;最后设计了符合优化趋势的非线性函数,将高斯平方距离转化为有利于神经网络学习的损失函数。实验结果表明,与IoU损失相比,所提方法在掩膜区域卷积神经网络、一阶全卷积目标检测器和自适应特征选择目标检测器上的平均精度均值分别提高了0.3%、1.1%和2.3%,证明了该方法能有效提升目标检测的性能,同时有利于高精度边界框的回归。展开更多
边界框回归分支是深度目标跟踪器的关键模块,其性能直接影响跟踪器的精度.评价精度的指标之一是交并比(Intersection over union,IoU).基于IoU的损失函数取代了l_(n)-norm损失成为目前主流的边界框回归损失函数,然而IoU损失函数存在2个...边界框回归分支是深度目标跟踪器的关键模块,其性能直接影响跟踪器的精度.评价精度的指标之一是交并比(Intersection over union,IoU).基于IoU的损失函数取代了l_(n)-norm损失成为目前主流的边界框回归损失函数,然而IoU损失函数存在2个固有缺陷:1)当预测框与真值框不相交时IoU为常量0,无法梯度下降更新边界框的参数;2)在IoU取得最优值时其梯度不存在,边界框很难收敛到IoU最优处.揭示了在回归过程中IoU最优的边界框各参数之间蕴含的定量关系,指出在边界框中心处于特定位置时存在多种尺寸不同的边界框使IoU损失最优的情况,这增加了边界框尺寸回归的不确定性.从优化两个统计分布之间散度的视角看待边界框回归问题,提出了光滑IoU(Smooth-IoU,SIoU)损失,即构造了在全局上光滑(即连续可微)且极值唯一的损失函数,该损失函数自然蕴含边界框各参数之间特定的最优关系,其唯一取极值的边界框可使IoU达到最优.光滑性确保了在全局上梯度存在使得边界框更容易回归到极值处,而极值唯一确保了在全局上可梯度下降更新参数,从而避开了IoU损失的固有缺陷.提出的光滑损失可以很容易取代IoU损失集成到现有的深度目标跟踪器上训练边界框回归,在LaSOT、GOT-10k、TrackingNet、OTB2015和VOT2018测试基准上所取得的结果,验证了光滑IoU损失的易用性和有效性.展开更多
As one of the primary computer vision problems, object detection aims to find and locate semantic objects in digital images. Different with object classification, which only recognizes an object to a certain class, ob...As one of the primary computer vision problems, object detection aims to find and locate semantic objects in digital images. Different with object classification, which only recognizes an object to a certain class, object detection also needs to extract accurate locations of objects. In the state-of-the-art object detection algorithms, bounding box regression plays a critical role in order to achieve high localization accuracy. Almost all the popular deep learning based object detection algorithms have utilized bounding box regression for fine tuning of object locations. However, while bounding box regression is widely used, there is few study focused on the underlying rationale, performance dependencies, and performance evaluation. In this paper, we proposed a dedicated deep neural network for bounding box regression, and presented several methods to improve its performance. Some ad hoc experiments are conducted to prove the effectiveness of the network. Also, we apply the network as an auxiliary module to the faster R-CNN algorithm and test them on some real-world images. Experiment results show certain performance improvements on detection accuracy in term of mean IOU.展开更多
文摘Camera-based object tracking systems in a given closed environment lack privacy and confidentiality.In this study,light detection and ranging(LiDAR)was applied to track objects similar to the camera tracking in a closed environment,guaranteeing privacy and confidentiality.The primary objective was to demonstrate the efficacy of the proposed technique through carefully designed experiments conducted using two scenarios.In Scenario I,the study illustrates the capability of the proposed technique to detect the locations of multiple objects positioned on a flat surface,achieved by analyzing LiDAR data collected from several locations within the closed environment.Scenario II demonstrates the effectiveness of the proposed technique in detecting multiple objects using LiDAR data obtained from a single,fixed location.Real-time experiments are conducted with human subjects navigating predefined paths.Three individuals move within an environment,while LiDAR,fixed at the center,dynamically tracks and identifies their locations at multiple instances.Results demonstrate that a single,strategically positioned LiDAR can adeptly detect objects in motion around it.Furthermore,this study provides a comparison of various regression techniques for predicting bounding box coordinates.Gaussian process regression(GPR),combined with particle swarm optimization(PSO)for prediction,achieves the lowest prediction mean square error of all the regression techniques examined at 0.01.Hyperparameter tuning of GPR using PSO significantly minimizes the regression error.Results of the experiment pave the way for its extension to various real-time applications such as crowd management in malls,surveillance systems,and various Internet of Things scenarios.
文摘在目标检测领域中,基于交并比(intersection over union, IoU)的系列损失函数存在一定的局限性,使得边界框回归的精度和稳定性有待进一步提升。为此提出了一种基于非线性高斯平方距离的边界框回归损失函数。首先综合考虑了边界框中重叠性、中心点距离和长宽比3个因素,将边界框建模为高斯分布;然后提出一种高斯平方距离来衡量概率分布之间的差距;最后设计了符合优化趋势的非线性函数,将高斯平方距离转化为有利于神经网络学习的损失函数。实验结果表明,与IoU损失相比,所提方法在掩膜区域卷积神经网络、一阶全卷积目标检测器和自适应特征选择目标检测器上的平均精度均值分别提高了0.3%、1.1%和2.3%,证明了该方法能有效提升目标检测的性能,同时有利于高精度边界框的回归。
文摘边界框回归分支是深度目标跟踪器的关键模块,其性能直接影响跟踪器的精度.评价精度的指标之一是交并比(Intersection over union,IoU).基于IoU的损失函数取代了l_(n)-norm损失成为目前主流的边界框回归损失函数,然而IoU损失函数存在2个固有缺陷:1)当预测框与真值框不相交时IoU为常量0,无法梯度下降更新边界框的参数;2)在IoU取得最优值时其梯度不存在,边界框很难收敛到IoU最优处.揭示了在回归过程中IoU最优的边界框各参数之间蕴含的定量关系,指出在边界框中心处于特定位置时存在多种尺寸不同的边界框使IoU损失最优的情况,这增加了边界框尺寸回归的不确定性.从优化两个统计分布之间散度的视角看待边界框回归问题,提出了光滑IoU(Smooth-IoU,SIoU)损失,即构造了在全局上光滑(即连续可微)且极值唯一的损失函数,该损失函数自然蕴含边界框各参数之间特定的最优关系,其唯一取极值的边界框可使IoU达到最优.光滑性确保了在全局上梯度存在使得边界框更容易回归到极值处,而极值唯一确保了在全局上可梯度下降更新参数,从而避开了IoU损失的固有缺陷.提出的光滑损失可以很容易取代IoU损失集成到现有的深度目标跟踪器上训练边界框回归,在LaSOT、GOT-10k、TrackingNet、OTB2015和VOT2018测试基准上所取得的结果,验证了光滑IoU损失的易用性和有效性.
文摘As one of the primary computer vision problems, object detection aims to find and locate semantic objects in digital images. Different with object classification, which only recognizes an object to a certain class, object detection also needs to extract accurate locations of objects. In the state-of-the-art object detection algorithms, bounding box regression plays a critical role in order to achieve high localization accuracy. Almost all the popular deep learning based object detection algorithms have utilized bounding box regression for fine tuning of object locations. However, while bounding box regression is widely used, there is few study focused on the underlying rationale, performance dependencies, and performance evaluation. In this paper, we proposed a dedicated deep neural network for bounding box regression, and presented several methods to improve its performance. Some ad hoc experiments are conducted to prove the effectiveness of the network. Also, we apply the network as an auxiliary module to the faster R-CNN algorithm and test them on some real-world images. Experiment results show certain performance improvements on detection accuracy in term of mean IOU.
文摘针对安全帽检测算法的快速且精准需求,提出了一种实时安全帽检测算法。首先,针对基于边界框回归损失函数容易出现梯度消失(Gradient Vanish)的问题,本文提出外接圆半径差(Circumcircle Radius Difference,CRD)损失函数;然后,针对复杂多尺度特征融合层制约检测速度的问题,提出了一种轻量化的小目标聚焦型(Focus on Small Object,FSO)特征融合层;最后本文结合YOLO网络、CRD和FSO形成YOLO-CRD-FSO(YCF)检测模型,实现实时安全帽检测。实验结果表明,在Jetson Xavier NX设备上检测分辨率为640×640的视频,YCF的检测速度达到43.4帧/秒,比当前最新锐的YOLO-V5模型的速度快了近2帧/秒,且均值平均精度提升了近1%。说明YCF检测模型综合优化了边界框回归损失函数和特征融合层,获得了良好的安全帽检测效果。
