融合迁移特征学习的生物启发网络用于可见光成像室内位置感知方法
With the rapid development of mobile communication technology in modern society, the demand for location services in complex indoor environments, such as large factories, shopping malls, and office buildings, has been growing rapidly. The current visible light positioning technology uses various sensors and hybrid complex algorithms to achieve positioning, which is difficult to operate and vulnerable to interference, resulting in unstable positioning accuracy of the system. The advantages of visible light communication include both lighting and communication, as well as stability and reliability. On this basis, to improve the accuracy and stability of visible light indoor positioning, a bio-inspired network integrating migration feature learning is proposed to achieve stable and high-precision indoor positioning in visible light imaging.
In this study, a visible-light indoor location method based on an image is proposed. The acquired image is first denoised to eliminate noise interference which has a significant impact on the extraction of the image depth features. Inaccurate feature extraction leads to poor positioning accuracy. An improved threshold denoising method is used to address the issue of signal loss caused by the oscillation of the threshold function. The adjustment function ensures good continuity of the signal and retains the original features of the image to the maximum extent based on image denoising. Second, the ResNet network is used to extract image depth features and establish a fingerprint database. The image depth features exhibit translation and rotation invariance. However, the ResNet network has deeper network layers than the traditional neural networks. Thus, residual learning is added to avoid a decrease in accuracy resulting from the increase in network layers. Finally, the BAS algorithm is used to optimize the connection weight matrix between the layers of the RBF neural network, improve the training speed and stability of the network, and determine the optimal weight between the layers of the network through back propagation for enhanced positioning accuracy.
In this study, we first build a positioning experimental platform (Fig. 4) consisting of a light environment that can simulate real indoor scenes to verify the applicability and effectiveness of the algorithm. The coordinate plate at the bottom of the experimental box is divided into several areas of equal size at an interval of 5 cm, and four LED light sources with the same size and power are installed on top of the experimental box to collect visible light images and extract depth features. Pictures are collected at three different heights by lifting and lowering the coordinate plate to establish a depth feature database for the collected pictures. The measured data are input into the neural network for training. The RBF neural network achieves 26983 target error iterations, the BAS optimized RBF neural network achieves 47352 iterations (Fig. 5), and the training speed is increased by approximately 40%. We randomly select 30 different coordinate points in the experimental box and collect the corresponding images without denoising for the positioning test. The average positioning error without denoising is 5.02 cm (Fig. 6), whereas, with denoising applied to the images collected at the same 30 points, the average positioning error is 4.26 cm (Fig. 7). The experiment shows that image denoising can effectively improve positioning accuracy. When compared to the RBF and back-propagation (BP)network algorithms, the BAS-RBF neural network algorithm provides significant improvements (Fig. 8) . Compared with the BP network algorithm, the confidence probability of a fixed error of less than 2 cm, 4 cm, and 6 cm increase by 9%, 11%, and 10%, respectively. The experimental results show that the performance of the RBF neural network optimized by BAS is better than those of the RBF and BP neural networks (Table 3). The average positioning error of the algorithm is 4.26 cm, which is 10.5% higher than that of the RBF neural network and 16.9% higher than that of the BP neural network.
This study proposes a visible light positioning technology for visual imaging that requires only images from indoor locations. Subsequently, the migration feature learning is used to extract the depth features of the denoised images to establish a database, which is brought into a neural network fused with biological algorithms for learning and training, with the goal of building a neural network training and testing model. Compared with the RBF and BP networks, this model can improve the positioning accuracy and training speed. In the actual measurement and positioning stage, 0.8 m×0.8 m×0.8 m physical model, the average positioning error of the prediction result is 4.26 cm; the probability of the prediction point error of less than 4 cm is 63.4% and the probability of the prediction point error of less than 6 cm is 78%. The positioning result is stable and reliable, providing a new feasible scheme for visible-light indoor positioning technology.
1 引言
随着当代社会移动通信技术的高速发展,大型工厂、商场及写字楼等复杂室内环境对定位服务的需求快速增长。目前应用最广泛的定位技术如全球定位系统(GPS)和北斗定位在室外环境中表现优异。但是,在室内环境中,在厚重障碍物、复杂环境和电磁屏蔽等因素影响下,常见的室内定位技术如红外定位、蓝牙定位等很难满足用户需求[1-2]。可见光定位技术作为一种新型的室内定位技术,相比于传统射频通信定位技术,具有兼顾照明和通信且稳定可靠的优点[3]。在新一代室内移动通信系统中,可见光室内定位技术已成为当下热门研究方向之一[4-5]。
近年来可见光定位技术已经有了一定的发展。常见的可见光定位系统主要使用光电检测器件采集信号强度值以及使用图像传感器采集图像,对信号及图像进行相应的处理以实现定位[6-8]。随着设备技术的完善和更多算法的使用,室内定位精度不断提高。例如:文献[9]提出一种非线性摄像机辅助接收信号强度算法,利用摄像机和光电二极管,结合视觉和强度信息实现定位;文献[10]针对室内光源信号互相叠加导致定位效果差的问题,提出利用码分多址对光源信息进行调制处理,对不同光源的信号进行分离,克服了码间干扰并提高了信号接收增益,定位误差达到6 cm;文献[11]提出了一种改进的隐马尔可夫模型算法,利用信号强度和节点间距建立模型,室内停车场的定位平均误差达到3.35 m;文献[12]提出了一种基于改进混合蝙蝠算法的可见光三维定位方法,在构造适应度函数时定义权重系数,并通过引入自适应搜索因子提高定位精度与速度,在1.5 m×1.2 m×2.0 m环境中平均定位误差为3.64 cm,定位时间为0.89 s;文献[13]利用粒子群算法优化神经网络,并压缩指纹数据库,缩短神经网络训练时间,网络对外部干扰具有很好的鲁棒性,提高了定位性能和精度;文献[14]利用图像传感器和加速传感器接受光源信息,基于两点光源进行粗定位,再利用2D图像坐标重建定位,在5 m×5 m×3 m的室内环境中定位误差和不超过2 cm,但该方法受接收端倾斜角度的影响,需根据不同环境不断改变内置传感器角度,实用性较差。
在传统室内定位技术不能满足人们日益增长的多元化定位需求的情况下,本文提出了一种基于图像的可见光室内定位方法,图像的深度特征具有平移旋转不变性,适用场景十分广阔。通过对采集的图像进行去噪处理并提取深度特征,建立图像深度特征指纹库,将深度特征融入到生物启发式神经网络模型中,可以更好地优化与筛选特征,实现稳定、高精度的室内定位。
2 可见光室内定位模型
为模拟真实的室内场景,本文搭建了0.8 m×0.8 m×0.8 m的可见光定位系统实验箱。在实验箱中,按照国际照度标准,基于光照度补偿原理,顶部布置
整个系统主要由离线建库、模型建立和实测定位三个部分构成。可见光成像室内定位的实验流程如
在离线建库阶段,采集图像并进行去噪处理,提取处理之后的图像的深度特征并建立指纹库。在模型建立阶段,搭建生物启发优化的径向基函数(RBF)神经网络模型,使用天牛须搜索(BAS)算法进行优化。在实测定位阶段,提取待测定位点的图像特征,将其作为输入神经元输入到训练好的定位模型中,设置目标误差,进行位置预测并获得预测坐标。
3 可见光成像的数据处理
3.1 可见光图像预处理
采集的图像常带有噪声干扰,噪声干扰对图像深度特征的提取有很大影响,特征提取不准确会导致定位精度较差[16]。采用的图像是手机摄像头拍摄的三原色(RGB)格式的图像,通常会受到传输设备本身以及外界环境的干扰,从而携带噪声。拍摄及传输过程中电路元器件自身固有的噪声及相互影响产生的噪声属于加性噪声。图像的噪声模型一般可表示为
式中:
式中:
3.2 基于迁移学习的图像特征提取
对图像进行去噪处理之后,就要提取图像深度特征并建立指纹库。训练的深度学习模型具有平移不变性,图像旋转、平移和缩放不会改变图像的深度特征。提取图像的深度特征需要搭建足够深度的网络。传统神经网络在训练过深的网络时,梯度会反向传到前面的层并不断相乘,产生梯度爆炸或梯度消失等问题,使神经网络的收敛更缓慢,准确率降低。
ResNet网络使用非常小的卷积滤波器架构来增加深度,相对于传统神经网络有更大的网络层数,同时为了避免网络层数的增加引起的准确度下降问题,增加了残差学习。残差网络在卷积层(conv)中引入了跳跃通道,经过卷积层与池化层(maxpool)的图像特征
4 基于生物启发优化神经网络的位置估计
4.1 BAS优化神经网络
BAS算法来源于天牛捕捉食物时的行为,仿照其觅食特性,从而对神经网络算法进行优化。原理是天牛额头上的两根触角对食物的气味具有很强的敏感性,例如当食物在右前方时,天牛两个触角会根据气味浓度差不断判断最优寻食路线,朝着气味浓度较大方向移动,直到抵达目标[18]。
RBF神经网络是多层前馈神经网络的一个特例,具有良好的逼近能力、非线性映射能力和分类能力等,可以从大量的数据中提取关键信息。使用径向基函数作为其激活函数,其公式为
式中:
在输入层输入采集坐标点图像的深度特征并在神经网络中进行训练,输出层对定位误差进行判断,使用BAS算法优化神经网络各层之间的连接权值矩阵,提高网络训练速度和稳定性,通过反向传播寻找网络各层之间的最优权值,从而提高定位精度。输出值是对三维坐标点的预测,需要定义三个神经元代表其输出坐标值。
4.2 创建数据集
在建立数据库时,在实验平台以等距间隔拍摄图片,提取特征建立数据库。假设样本总数为
式中:
式中:
该矩阵表示
式中:
式中:
5 定位实验与数据分析
5.1 实验环境及参数设置
论文搭建了
在实验箱底部的坐标板上以5 cm为间隔将坐标板划分为大小相等的若干区域,并在实验箱顶部安装4个功率相同的LED光源,用来采集可见光图像并提取深度特征;通过升降坐标板,在不同高度(0、0.11、0.19 m)处采集图片,利用采集的图片的深度特征建立数据库。将实测的1684组数据输入到神经网络中进行训练。在划分的区域内采集图像,提取每个点处采集图像的深度特征并建立数据库。
表 1. 实验环境参数
Table 1. Experimental environment parameters
|
5.2 基于BAS的网络训练
将RBF网络训练次数设置为1000,学习速率为0.015,输入层节点数为256,输出层节点数为3,设置训练网络出口,当误差小于0.04 m时,保存神经网络并退出训练。在BAS参数设置中,初始搜索步长为100,步长调整比例为0.5,触须间距为5,变量维数为2。对数据集进行训练,BAS优化后的RBF(BAS-RBF)神经网络与优化前的RBF神经网络的训练速度对比如
图 5. BAS优化前后RBF神经网络的训练速度
Fig. 5. Training speeds of RBF neural networks before and after BAS optimization
5.3 实验结果及分析
随机在实验箱内选取30个不同坐标点,采集对应图像但不对其进行去噪处理,定位测试结果如
图 6. 图像去噪前实测的定位误差分布
Fig. 6. Distribution of measured positioning error before image denoising
图 7. 图像去噪后实测的定位误差分布
Fig. 7. Distribution of measured positioning error after image denoising
图像去噪后得到的平均定位误差为4.26 cm,实验表明,对图像进行去噪能有效地提高定位精度。在图像去噪后的实测定位误差结果中选取10组位置数据,对真实坐标与预测坐标进行比较,结果如
表 2. 部分位置坐标的比较
Table 2. Comparison of partial position coordinates
|
对BAS-RBF神经网络算法、RBF神经网络算法、反向传播(BP)神经网络算法进行比较,得到的累计误差分布如
由
表 3. 三种算法的定位误差比较
Table 3. Comparison of positioning errors of three algorithms
|
由实验结果可知,经BAS优化的RBF神经网络算法的性能均高于RBF神经网络和BP神经网络。本文所提算法的平均定位误差为4.26 cm,相比RBF神经网络定位精度提高了10.5%,相比BP神经网络提高了16.9%。
6 结论
针对现行室内可见光定位技术受环境影响大、实际操作难度高导致的定位不稳定的问题,提出了视觉成像的可见光定位技术。只需在室内位置采集图像,再使用迁移特征学习提取去噪处理的图像的深度特征,建立数据库,并将其代入到融合生物算法的神经网络中进行学习训练,搭建神经网络训练与测试模型。在定位精度和训练速度上,该模型相比RBF网络和BP网络均有提升。在实测定位阶段,搭建了0.8 m×0.8 m×0.8 m的实物模型进行验证,预测结果的平均定位误差为4.26 cm,误差小于4 cm的预测点数量占总坐标点数量的63.4%,误差小于6 cm的预测点数量占总坐标点数量的78%。定位结果稳定可靠,所提方法为可见光室内定位技术提供了新的可行方案。
[1] 赵黎, 韩中达, 张峰. 基于神经网络的可见光室内立体定位研究[J]. 中国激光, 2021, 48(7): 0706004.
[3] Wang C Y, Wang L, Chi X F, et al. The research of indoor positioning based on visible light communication[J]. China Communications, 2015, 12(8): 85-92.
[4] 白勃, 田骅, 李晓波. 基于发光二极管光源的相机定标无线定位算法[J]. 中国激光, 2015, 42(2): 0205003.
[5] 亢烨, 柯熙政. 可见光通信中的多维编码[J]. 中国激光, 2015, 42(2): 0205001.
[6] 贾兰芳, 周小芳, 吕丽君. 基于光信号强度检测法的可见光通信定位方法研究[J]. 激光杂志, 2021, 42(5): 97-101.
Jia L F, Zhou X F, Lv L J. Research on visible light communication location method based on opticalsignal intensity detection[J]. Laser Journal, 2021, 42(5): 97-101.
[7] Li Y W, Ghassemlooy Z, Tang X, et al. A VLC smartphone camera based indoor positioning system[J]. IEEE Photonics Technology Letters, 2018, 30(13): 1171-1174.
[8] 张志涛, 郭彩丽, 柏琳, 等. PD辅助的P3P可见光定位算法[J]. 北京邮电大学学报, 2022, 45(1): 75-81.
Zhang Z T, Guo C L, Bai L, et al. Perspective-three-point visible light localization algorithm assisted by photo-diode[J]. Journal of Beijing University of Posts and Telecommunications, 2022, 45(1): 75-81.
[9] BaiB, SuB, ChenN, et al. A high precision positioning method based on high speed camera and visible light[C]∥ 2017 Asia Communications and Photonics Conference (ACP), November 10-13, 2017, Guangzhou, China. New York: IEEE Press, 2017.
[10] 关伟鹏, 吴玉香, 文尚胜, 等. 基于CDMA调制的可见光通信室内定位技术的研究[J]. 光学学报, 2016, 36(11): 1106006.
[11] 陈勇, 巫杰, 刘焕淋, 等. 基于隐马尔可夫模型的可见光与惯导融合室内定位系统[J]. 中国激光, 2020, 47(12): 1206001.
[12] Chen Y, Zheng H, Liu H L, et al. Indoor high precision three-dimensional positioning system based on visible light communication using improved hybrid bat algorithm[J]. IEEE Photonics Journal, 2020, 12(5): 6802513.
[13] Zhang M Q, Li F J, Guan W P, et al. A three-dimensional indoor positioning technique based on visible light communication using chaotic particle swarm optimization algorithm[J]. Optik, 2018, 165: 54-73.
[14] 车平, 冯莉芳. 基于室内可见光的两点光源定位技术研究[J]. 光学技术, 2018, 44(3): 257-262.
[15] 刘智港, 赵黎, 朱彤, 等. 兼顾照明的室内可见光定位系统研究[J]. 激光技术, 2018, 42(5): 646-650.
[16] 李敏. 基于新阈值函数的小波图像去噪算法研究[J]. 控制工程, 2021, 28(12): 2360-2365.
Li M. Research on wavelet image denoising algorithm based on new threshold function[J]. Control Engineering of China, 2021, 28(12): 2360-2365.
[17] 沙宇洋, 席丽霞, 张晓光, 等. 基于小波阈值去噪的偏振模色散测量[J]. 中国激光, 2018, 45(11): 1106006.
[18] Khan A T, Cao X W, Li S, et al. Fraud detection in publicly traded US firms using Beetle Antennae Search: a machine learning approach[J]. Expert Systems with Applications, 2022, 191: 116148.
Article Outline
孟祥艳, 张欣, 张峰, 赵黎, 李帅. 融合迁移特征学习的生物启发网络用于可见光成像室内位置感知方法[J]. 中国激光, 2023, 50(10): 1006007. Xiangyan Meng, Xin Zhang, Feng Zhang, Li Zhao, Shuai Li. Bioheuristic Network Based on Migration Feature Learning for Indoor Location Awareness in Visible Light Imaging[J]. Chinese Journal of Lasers, 2023, 50(10): 1006007.