技术领域Technical Field
本发明涉及安全监控技术领域，特别是涉及一种基于人脸识别技术的社区儿童安全追踪系统。The present invention relates to the technical field of safety monitoring, and in particular to a community child safety tracking system based on face recognition technology.
背景技术Background Art
随着城市化进程的加快，社区居住人口密度逐渐增加，儿童的安全问题成为家长和社区管理者关注的焦点。目前，传统的监控系统主要依靠视频监控和人工巡逻，存在监控盲区大、实时性差、效率低等问题，难以满足儿童安全管理的需求。人脸识别技术作为一种生物特征识别技术，具有识别准确、反应迅速、应用范围广等优点，逐渐被应用于各类安全监控系统中。因考虑到儿童安全事件可能发生的时间段为7*24小时，故本发明引入了针对夜视红外场景下儿童人脸识别技术。With the acceleration of urbanization, the population density of communities has gradually increased, and the safety of children has become the focus of attention of parents and community managers. At present, traditional monitoring systems mainly rely on video monitoring and manual patrols, which have problems such as large monitoring blind spots, poor real-time performance, and low efficiency, making it difficult to meet the needs of child safety management. As a biometric recognition technology, face recognition technology has the advantages of accurate recognition, rapid response, and a wide range of applications, and has gradually been applied to various security monitoring systems. Considering that the time period in which child safety incidents may occur is 7*24 hours, the present invention introduces a child face recognition technology for night vision infrared scenes.
发明内容Summary of the invention
本发明的目的在于提供一种基于人脸识别技术的社区儿童安全追踪系统，以解决现有技术中存在的监控盲区大、实时性差、效率低等问题，实现对儿童的实时、精确、安全的追踪和监控，因儿童人脸特征点会随着其成长而发生变化，故引入特征修正功能。因有夜视场景，故引入夜视场景下的儿童人脸识别技术。The purpose of the present invention is to provide a community child safety tracking system based on face recognition technology to solve the problems of large monitoring blind spots, poor real-time performance, and low efficiency in the prior art, and to achieve real-time, accurate, and safe tracking and monitoring of children. Because the facial feature points of children will change as they grow, a feature correction function is introduced. Because there are night vision scenes, child face recognition technology in night vision scenes is introduced.
为了实现上述目的，本发明提供了一种基于人脸识别的社区儿童安全追踪系统，该系统包括：数据库模块、监控模块、人脸识别模块、显示模块、报警模块、数据分析模块、人脸特征底库修正模块；In order to achieve the above-mentioned object, the present invention provides a community child safety tracking system based on face recognition, which comprises: a database module, a monitoring module, a face recognition module, a display module, an alarm module, a data analysis module, and a face feature database correction module;
数据库模块存储儿童的面部特征信息、家长联系方式以及儿童的日常活动轨迹；The database module stores the child’s facial feature information, parent contact information, and the child’s daily activity track;
监控模块包括分布在社区的监控摄像头，设置在社区的关键区域，用于实时捕捉儿童的活动图像，并传送至人脸识别模块进行处理；The monitoring module includes surveillance cameras distributed throughout the community, which are set up in key areas of the community to capture children's activity images in real time and transmit them to the face recognition module for processing;
人脸识别模块：用于捕捉和识别进入社区的儿童面部特征信息，并与数据库中的预存信息进行比对，确认儿童身份；Face recognition module: used to capture and identify the facial features of children entering the community, and compare them with the pre-stored information in the database to confirm the identity of the children;
显示模块：用于显示儿童的实时位置和活动轨迹，便于家长和社区管理人员进行监控；显示屏通过WebSocket协议与系统服务器通信，实时刷新数据，延迟小于1秒；Display module: used to display the real-time location and activity trajectory of children, which is convenient for parents and community managers to monitor; the display screen communicates with the system server through the WebSocket protocol, refreshes data in real time, and the delay is less than 1 second;
报警模块：当系统检测到儿童进入危险区域或长期停留在异常位置时，报警模块及时通过短信网关和手机App自动向家长和社区管理人员发送警报信息；Alarm module: When the system detects that a child has entered a dangerous area or has stayed in an abnormal location for a long time, the alarm module will automatically send an alarm message to parents and community managers through the SMS gateway and mobile phone app;
数据分析模块对儿童的日常活动数据进行统计分析。生成的分析报告包括活动频率、活动时长、活动区域分布等信息，并能够识别异常行为模式，提供潜在风险的预警报告；The data analysis module performs statistical analysis on children's daily activity data. The generated analysis report includes information such as activity frequency, activity duration, and activity area distribution. It can also identify abnormal behavior patterns and provide early warning reports of potential risks.
人脸特征底库修正模块定期对数据库中的面部特征信息进行更新和修正，以确保系统识别准确性。The facial feature database correction module regularly updates and corrects the facial feature information in the database to ensure the accuracy of system recognition.
进一步的，所述数据库模块采用MySQL数据库，数据库中存储的面部特征数据为256维的特征向量，通过Dlib库生成。Furthermore, the database module adopts a MySQL database, and the facial feature data stored in the database is a 256-dimensional feature vector generated by a Dlib library.
进一步的，所述社区的关键区域包括出入口、游乐场、停车场；摄像头支持25帧每秒的实时视频捕捉，确保视频数据的清晰度和流畅度。Furthermore, key areas of the community include entrances and exits, playgrounds, and parking lots; the camera supports real-time video capture at 25 frames per second to ensure the clarity and smoothness of video data.
进一步的，所述人脸识别模块包括数据预处理模块、人脸识别模型；Furthermore, the face recognition module includes a data preprocessing module and a face recognition model;
所述数据预处理模块的工作流程具体如下：The workflow of the data preprocessing module is as follows:
步骤A1：从监控模块中实时获取原始人脸图像；Step A1: obtaining the original face image from the monitoring module in real time;
步骤A2：将捕捉到的RGB彩色图像像素值归一化；Step A2: normalizing the pixel values of the captured RGB color image;
步骤A3：使用Dlib库中的人脸检测工具，自动检测并对齐面部关键点；Step A3: Use the face detection tool in the Dlib library to automatically detect and align facial key points;
步骤A4：使用非对称直方图均衡化进行图像增强；Step A4: Image enhancement using asymmetric histogram equalization;
人脸识别模型具体为ResNet-50网络模型；经过预处理的人脸特征输入至ResNet-50网络，输出的特征向量输入milvus特征向量数据库，数据库通过内部余弦相似度算法对比得到高于相似度阈值的人脸特征向量，作为人脸识别结果；The face recognition model is specifically a ResNet-50 network model. The preprocessed face features are input into the ResNet-50 network, and the output feature vectors are input into the milvus feature vector database. The database is compared through the internal cosine similarity algorithm to obtain the face feature vectors above the similarity threshold as the face recognition results.
进一步的，所述数据分析模块使用Hadoop大数据平台进行数据处理和分析。Furthermore, the data analysis module uses the Hadoop big data platform to perform data processing and analysis.
进一步的，人脸特征底库修正模块的修正过程具体为：每个月从识别记录中挑选出每一位儿童图片质量最高、对比相似度最高并且超过最低阈值的人脸照片，用来替换成该位儿童最新的人脸底图；若超过6个月系统没有通过自动修正更新底图，则产生报警，提醒重新进行人工录入。Furthermore, the correction process of the facial feature base database correction module is as follows: every month, the facial photo of each child with the highest image quality, the highest comparison similarity and exceeding the minimum threshold is selected from the recognition records to replace the latest facial base map of the child; if the system fails to update the base map through automatic correction for more than 6 months, an alarm is generated to remind people to re-enter the map manually.
所述人脸识别模型ResNet-50网络的训练方法如下：The training method of the face recognition model ResNet-50 network is as follows:
步骤B1：数据集准备；获取儿童面部图像数据集，数据集覆盖不同年龄段、性别、国家的儿童面部图像；Step B1: Dataset preparation: obtain a children's facial image dataset, which covers children's facial images of different age groups, genders, and countries;
步骤B2：数据预处理，并使用非对称直方图均衡化对夜视图片进行图像增强；然后将数据集划分为训练集、验证集和测试集；Step B2: Data preprocessing and image enhancement of night vision images using asymmetric histogram equalization; then the dataset is divided into training set, validation set and test set;
步骤B3：模型训练；采用Adam优化算法进行梯度下降；Step B3: Model training; using Adam optimization algorithm for gradient descent;
步骤B4：模型优化与评估；使用准确率、F1-score、ROC曲线评价指标衡量模型在验证集上的表现，得到训练好的人脸识别模型，应用于人脸识别模块中。Step B4: Model optimization and evaluation; use accuracy, F1-score, and ROC curve evaluation indicators to measure the performance of the model on the validation set, obtain a trained face recognition model, and apply it to the face recognition module.
本发明的有益效果如下：The beneficial effects of the present invention are as follows:
本发明基于人脸识别技术，实现了对儿童的实时、精确、安全的追踪和监控，解决了传统监控系统监控盲区大、实时性差、效率低的问题，有效提升了社区儿童的安全保障水平；在传统监控图像中容易出现过曝或者太暗的问题，本发明中的数据分析模块使用非对称均衡化方法使图像的暗部区域被增强，亮部保持适当亮度，从而整体提升了图像的细节和可见性，让人脸图像更加清晰，提高识别正确率。The present invention is based on face recognition technology, realizes real-time, accurate and safe tracking and monitoring of children, solves the problems of large blind spots, poor real-time performance and low efficiency of traditional monitoring systems, and effectively improves the safety level of children in the community; in traditional monitoring images, overexposure or too dark problems are prone to occur. The data analysis module in the present invention uses an asymmetric equalization method to enhance the dark areas of the image and maintain appropriate brightness in the bright areas, thereby improving the details and visibility of the image as a whole, making the face image clearer and improving the recognition accuracy.
附图说明BRIEF DESCRIPTION OF THE DRAWINGS
图1为本发明基于人脸识别的社区儿童安全追踪系统的整体架构图。FIG1 is an overall architecture diagram of a community child safety tracking system based on face recognition according to the present invention.
具体实施方式DETAILED DESCRIPTION
在某社区内部署了一套基于人脸识别的儿童安全追踪系统，如图1所示，该系统包括数据库模块、监控模块、人脸识别模块、显示模块、报警模块、数据分析模块以及人脸特征底库修正模块；当儿童进入社区时，监控模块捕捉到其面部图像并传送至人脸识别模块。人脸识别模块将捕捉到的图像与数据库中的预存信息进行比对，确认儿童身份，系统实时跟踪儿童的活动轨迹，并将数据传送至数据库模块进行存储，当系统检测到儿童进入危险区域或长期停留在异常位置时，报警模块自动发送警报信息至家长和社区管理人员；显示模块实时显示儿童的位置和活动轨迹，便于监控；数据分析模块对儿童的日常活动数据进行分析，提供行为模式和潜在风险的预警报告。A child safety tracking system based on face recognition has been deployed in a community. As shown in Figure 1, the system includes a database module, a monitoring module, a face recognition module, a display module, an alarm module, a data analysis module, and a face feature database correction module. When a child enters the community, the monitoring module captures his or her facial image and transmits it to the face recognition module. The face recognition module compares the captured image with the pre-stored information in the database to confirm the child's identity. The system tracks the child's activity trajectory in real time and transmits the data to the database module for storage. When the system detects that a child enters a dangerous area or stays in an abnormal position for a long time, the alarm module automatically sends an alarm message to the parents and community managers. The display module displays the child's location and activity trajectory in real time for easy monitoring. The data analysis module analyzes the child's daily activity data and provides early warning reports on behavior patterns and potential risks.
接下来对各模块进行进一步阐述。Next, each module is further explained.
数据库模块采用MySQL数据库，存储儿童的面部特征信息、家长联系方式以及儿童的日常活动轨迹，数据库中存储的面部特征数据为256维的特征向量，通过Dlib库生成。The database module uses MySQL database to store children's facial feature information, parents' contact information, and children's daily activity trajectories. The facial feature data stored in the database is a 256-dimensional feature vector generated by the Dlib library.
监控模块包括分布在社区的若干个监控摄像头，设置在社区的各个关键区域，用于实时捕捉儿童的活动图像，并传送至人脸识别模块进行处理。具体的，分布在社区的主要出入口、游乐场、停车场等关键区域。摄像头型号为海康DS-2XA3T46EF-LS，具有400万像素，支持25帧每秒的实时视频捕捉，确保视频数据的清晰度和流畅度。The monitoring module includes several surveillance cameras distributed in the community, which are set up in various key areas of the community to capture children's activities in real time and transmit them to the face recognition module for processing. Specifically, they are distributed in key areas such as the main entrances and exits of the community, playgrounds, and parking lots. The camera model is Hikvision DS-2XA3T46EF-LS, with 4 million pixels and support for real-time video capture at 25 frames per second, ensuring the clarity and smoothness of video data.
人脸识别模块：用于捕捉和识别进入社区的儿童面部特征信息，并与数据库中的预存信息进行比对，确认儿童身份。Face recognition module: used to capture and identify the facial features of children entering the community, and compare it with the pre-stored information in the database to confirm the identity of the children.
显示模块：用于显示儿童的实时位置和活动轨迹，便于家长和社区管理人员进行监控。显示屏通过WebSocket协议与系统服务器通信，实时刷新数据，延迟小于1秒。Display module: used to display the real-time location and activity trajectory of children, which is convenient for parents and community managers to monitor. The display screen communicates with the system server through the WebSocket protocol and refreshes data in real time with a delay of less than 1 second.
报警模块：当系统检测到儿童进入危险区域（如停车场）或长期停留在异常位置（如超过15分钟）时，报警模块及时通过短信网关和手机App自动向家长和社区管理人员发送警报信息，响应时间小于5秒，确保及时报警。Alarm module: When the system detects that a child enters a dangerous area (such as a parking lot) or stays in an abnormal location for a long time (such as more than 15 minutes), the alarm module will automatically send alarm information to parents and community managers through the SMS gateway and mobile phone App. The response time is less than 5 seconds to ensure timely alarm.
数据分析模块使用Hadoop大数据平台进行数据处理和分析，对儿童的日常活动数据进行统计分析。生成的分析报告包括活动频率、活动时长、活动区域分布等信息，并能够识别异常行为模式，提供潜在风险的预警报告。The data analysis module uses the Hadoop big data platform to process and analyze data and conduct statistical analysis on children's daily activity data. The generated analysis report includes information such as activity frequency, activity duration, and activity area distribution, and can identify abnormal behavior patterns and provide early warning reports of potential risks.
人脸特征底库修正模块定期对数据库中的面部特征信息进行更新和修正，以确保系统识别准确性。具体的，系统每个月从识别记录中挑选出每一位儿童图片质量最高、对比相似度最高并且超过最低阈值的人脸照片，用来替换成该位儿童最新的人脸底图。基于历史数据的变化，系统能够自动调整面部特征数据，以适应儿童面部特征的成长变化，确保长期识别准确性。若超过6个月系统没有通过自动修正功能更新底图，则产生报警，提醒重新进行人工录入。The facial feature base database correction module regularly updates and corrects the facial feature information in the database to ensure the accuracy of system recognition. Specifically, the system selects the facial photos of each child with the highest image quality, the highest comparison similarity and exceeding the minimum threshold from the recognition records every month to replace them with the latest facial base map of the child. Based on changes in historical data, the system can automatically adjust facial feature data to adapt to the growth and changes of children's facial features to ensure long-term recognition accuracy. If the system has not updated the base map through the automatic correction function for more than 6 months, an alarm will be generated to remind you to re-enter the data manually.
具体的，人脸识别模块包括数据预处理模块、人脸识别模型；数据预处理模块的工作流程具体如下：Specifically, the face recognition module includes a data preprocessing module and a face recognition model; the workflow of the data preprocessing module is as follows:
步骤A1：从监控模块中实时获取原始人脸图像；Step A1: obtaining the original face image from the monitoring module in real time;
步骤A2：灰度归一化：将捕捉到的RGB彩色图像像素值归一化到[0, 1]范围；灰度转换公式为：Step A2: Grayscale normalization: Normalize the captured RGB color image pixel values to the range [0, 1]. The grayscale conversion formula is:
其中，R,G,B分别为红、绿、蓝通道的像素值，Y为转换后的灰度值。Among them, R, G, B are the pixel values of the red, green, and blue channels respectively, and Y is the grayscale value after conversion.
步骤A3：图像面部对齐：使用Dlib库中的人脸检测工具，自动检测并对齐面部关键点，确保面部在图像中的位置和尺度一致；Step A3: Image face alignment: Use the face detection tool in the Dlib library to automatically detect and align facial key points to ensure that the position and scale of the face in the image are consistent;
步骤A4：使用非对称直方图均衡化对图像进行增强；Step A4: using asymmetric histogram equalization to enhance the image;
步骤A4.1：双区段非对称划分：将图像的直方图分为两个部分，分别是高亮度区和低亮度区，并对两个区域分别采用不同的均衡化策略；Step A4.1: Dual-segment asymmetric partitioning: the histogram of the image is divided into two parts, namely, a high brightness area and a low brightness area, and different equalization strategies are used for the two areas respectively;
每个像素都有一个对应的坐标 (x, y)，并且该像素的灰度值表示该点的亮度，取值（0，255）。低亮度区：对低亮度部分进行细致的均衡化，增强暗部细节，同时限制亮度值的提升，防止过度曝光，对于低亮度区L(x,y)，L(x,y)取值在(0,95)；高亮度区：对于高亮度部分，采用较缓和的均衡化策略，避免亮部区域过度增强，导致细节丢失，对于低亮度区H(x,y)，H(x,y)取值在(183,255)。Each pixel has a corresponding coordinate (x, y), and the grayscale value of the pixel represents the brightness of the point, with a value of (0, 255). Low brightness area: Perform detailed equalization on the low brightness part to enhance the dark details, while limiting the increase in brightness to prevent overexposure. For the low brightness area L(x,y), L(x,y) takes the value of (0,95); High brightness area: For the high brightness part, a more moderate equalization strategy is adopted to avoid over-enhancement of the bright area, resulting in loss of details. For the low brightness area H(x,y), H(x,y) takes the value of (183,255).
对低亮度区L(x,y),对灰度值执行局部直方图均衡化：For the low brightness area L(x,y), perform local histogram equalization on the grayscale value:
其中，为暗部增强系数，用来调整暗部细节，表示图像中低亮度区域的像素值，L表示图像中低亮度区；表示均衡化后的低亮度区域的像素值。in, It is the dark enhancement coefficient, which is used to adjust the dark details. Indicates the pixel value of the low brightness area in the image, L represents the low brightness area in the image; Indicates the pixel value of the low-brightness area after equalization.
对高亮度区域应用平缓的均衡化策略，避免亮部过曝；高亮度区域的平缓均衡化通过以下公式计算：Apply a gentle equalization strategy to high-brightness areas to avoid overexposure of bright areas; the gentle equalization of high-brightness areas is calculated by the following formula:
其中，为亮部限制系数，防止亮部过曝，表示图像中高亮度区域的像素值，H表示图像中高亮度区；表示均衡化后的高亮度区域的像素值；in, It is the highlight limit coefficient to prevent overexposure of highlights. Indicates the pixel value of the high brightness area in the image, H represents the high brightness area in the image; Represents the pixel value of the high brightness area after equalization;
步骤A4.2：自适应权重分配：根据图像中高亮度区和低亮度区这两个区域来动态调整权重。对于灰度分布较为集中的区域，降低均衡化的力度，以保持图像的自然性；对于灰度分布广泛的区域，则增加均衡化的强度，提升局部对比度。Step A4.2: Adaptive weight allocation: Dynamically adjust the weights based on the high brightness area and low brightness area in the image. For areas with concentrated grayscale distribution, reduce the intensity of equalization to maintain the naturalness of the image; for areas with widespread grayscale distribution, increase the intensity of equalization to improve local contrast.
灰度方差计算：Grayscale variance calculation:
其中，n是区域内的像素数，区域包括高亮度区H(x,y)和低亮度区L(x,y)，是位于处的灰度值，是该区域的平均灰度值。Where n is the area The number of pixels within the area Including high brightness area H(x,y) and low brightness area L(x,y), is located in The gray value at is the average gray value of the area.
局部区域权重由灰度分布的方差 决定：Local area weight The variance of the grayscale distribution Decide:
方差较大的区域得到较大的权重，方差较小的区域则降低处理强度。Regions with larger variances are given larger weights, while regions with smaller variances are treated less strongly.
应用于非对称直方图均衡化中，公式表示如下：Applied to asymmetric histogram equalization, the formula is as follows:
低亮度区的处理，T1取值95：Processing of low brightness areas , T1 takes the value 95:
其中，是低亮度区的动态调整权重。in, It is the dynamic adjustment weight of low brightness area.
高亮度区的处理,T2取值193：High brightness area processing ,T2 takes the value 193:
其中，是高亮度区的动态调整权重，和为低亮度和高亮度阈值。in, is the dynamic adjustment weight of the high brightness area, and are the low brightness and high brightness thresholds.
低亮度区域如果灰度分布较为集中（即区域内的像素亮度变化不大），则均衡化力度减弱，避免过度增强暗部的噪声。高亮度区域如果灰度分布较广，则增加均衡化的力度，提升该区域的对比度，使其细节更加清晰。通过灰度方差计算，动态调整高亮度区和低亮度区的均衡化权重，使得图像处理更加细腻。这种方法特别适合处理亮度差异较大的场景，如夜间图像、红外图像等，能够在增强图像细节的同时，保持自然感。If the grayscale distribution in the low-brightness area is relatively concentrated (i.e., the pixel brightness in the area does not change much), the equalization strength is weakened to avoid excessive enhancement of the noise in the dark area. If the grayscale distribution in the high-brightness area is relatively wide, the equalization strength is increased to improve the contrast of the area and make its details clearer. By calculating the grayscale variance, the equalization weights of the high-brightness area and the low-brightness area are dynamically adjusted to make the image processing more delicate. This method is particularly suitable for processing scenes with large brightness differences, such as night images and infrared images, and can enhance image details while maintaining a natural feel.
步骤A4.3：基于边缘检测的非对称均衡化：引入边缘检测机制，在直方图均衡化之前对图像的边缘区域进行预处理。通过对边缘区域进行高精度均衡化，能够保留图像中的关键细节和轮廓信息，防止传统直方图均衡化对边缘的过度平滑：Step A4.3: Asymmetric equalization based on edge detection: Introduce edge detection mechanism to preprocess the edge area of the image before histogram equalization. By performing high-precision equalization on the edge area, key details and contour information in the image can be retained, preventing the traditional histogram equalization from over-smoothing the edge:
对边缘区域应用更细化的非对称均衡化策略，而对非边缘区域应用较缓和的增强。A more refined asymmetric equalization strategy is applied to edge areas, while a milder enhancement is applied to non-edge areas.
边缘区域的非对称均衡化： 对于检测到的边缘区域，应用更细化的非对称均衡化策略，特别是增强低亮度部分的细节。边缘区域的均衡化公式可以表示为：Asymmetric equalization of edge areas: For detected edge areas, a more refined asymmetric equalization strategy is applied, especially to enhance the details of low-brightness parts. The equalization formula for edge areas can be expressed as:
表示原始图像中像素点的灰度值，表示边缘区域中经过均衡化处理后的像素值，和为低亮度和高亮度阈值，用于定义增强的区间；和是增强系数，用于控制边缘区域的增强力度。 Represents the grayscale value of the pixel in the original image. Represents the pixel value after equalization in the edge area. and are low brightness and high brightness thresholds, used to define the enhancement interval; and is the enhancement factor, which is used to control the enhancement strength of the edge area.
步骤A4.4：自适应动态范围压缩：为避免在高亮度图像中出现过度曝光和在低亮度图像中出现暗部丢失的问题，采用自适应动态范围压缩方法。根据图像的局部亮度直方图对不同区域的动态范围进行压缩，保证整个图像的亮度分布更加均匀。动态范围的压缩程度由图像的全局对比度自动调整。Step A4.4: Adaptive dynamic range compression: To avoid overexposure in high-brightness images and loss of dark areas in low-brightness images, an adaptive dynamic range compression method is used. The dynamic range of different regions is compressed according to the local brightness histogram of the image to ensure a more uniform brightness distribution of the entire image. The degree of dynamic range compression is automatically adjusted by the global contrast of the image.
其中 C为对比度调整参数，I(x, y) 为输入图像像素值， 为压缩后的输出值。Where C is the contrast adjustment parameter, I(x, y) is the input image pixel value, is the compressed output value.
经过非对称直方图均衡化处理后，图像的暗部区域被增强，亮部保持适当亮度，从而整体提升了图像的细节和可见性；以具体的图像为例，原始图像的平均灰度值为70，经过均衡化处理后，平均灰度值上升到120，标准差从15提升至40，表明图像对比度显著提高。After asymmetric histogram equalization processing, the dark areas of the image are enhanced and the bright areas maintain appropriate brightness, thereby improving the details and visibility of the image as a whole; taking a specific image as an example, the average grayscale value of the original image is 70. After equalization processing, the average grayscale value rises to 120, and the standard deviation increases from 15 to 40, indicating that the image contrast is significantly improved.
人脸识别模型具体为ResNet-50网络模型；ResNet-50网络是一种50层的残差神经网络，通过引入残差连接解决了深层网络训练过程中梯度消失的问题；具体的，ResNet-50网络包含一系列卷积层和残差块，每个残差块包含多个卷积层、批量归一化（BatchNormalization）和ReLU激活函数，目的是提取图像的空间特征，最后经过全连接层输出一个256维的特征向量，作为每张人脸的特征表示，这个特征向量代表了人脸的关键信息，将这个特征向量输入milvus向量数据库，系统中通过milvus数据库进行余弦相似度计算对比，得到相似度最高的人员id与对比相似度，经过识别结果与真人比对，发现当相似度阈值设置为0.85时准确率可达99.998%。The face recognition model is specifically the ResNet-50 network model; the ResNet-50 network is a 50-layer residual neural network that solves the problem of gradient disappearance during deep network training by introducing residual connections; specifically, the ResNet-50 network contains a series of convolutional layers and residual blocks, each residual block contains multiple convolutional layers, batch normalization (BatchNormalization) and ReLU activation functions, the purpose of which is to extract the spatial features of the image, and finally output a 256-dimensional feature vector through the fully connected layer as the feature representation of each face. This feature vector represents the key information of the face, and this feature vector is input into the milvus vector database. The system uses the milvus database to perform cosine similarity calculation and comparison to obtain the person ID with the highest similarity and the comparison similarity. After comparing the recognition results with the real person, it is found that when the similarity threshold is set to 0.85, the accuracy can reach 99.998%.
具体的，人脸识别模型ResNet-50网络的训练方法如下：Specifically, the training method of the face recognition model ResNet-50 network is as follows:
步骤B1：数据集准备；获取儿童面部图像数据集，数据集覆盖了不同年龄段、性别、种族的儿童面部图像；Step B1: Dataset preparation: Obtain a children's facial image dataset, which covers facial images of children of different ages, genders, and races;
步骤B2：数据预处理，包括数据增强、图像面部对齐、灰度归一化、夜视图像增强；通过数据增强技术扩充数据集，具体操作包括旋转、翻转、缩放、平移、加噪等，以模拟各种可能的场景变化，提高模型的泛化能力。因存在夜视场景下的人脸识别，故本模型训练的数据集中有约30%的图片为夜视场景下抓取到的人脸图片。夜视图像增强具体使用非对称直方图均衡化对夜视图片进行图像增强；然后将数据集划分为训练集（80%）、验证集（10%）和测试集（10%），以评估模型的训练效果和泛化能力。Step B2: Data preprocessing, including data enhancement, image face alignment, grayscale normalization, and night vision image enhancement; expand the data set through data enhancement technology, including rotation, flipping, scaling, translation, noise addition, etc., to simulate various possible scene changes and improve the generalization ability of the model. Because of the existence of face recognition in night vision scenes, about 30% of the images in the data set trained by this model are face images captured in night vision scenes. Night vision image enhancement specifically uses asymmetric histogram equalization to enhance night vision images; then the data set is divided into a training set (80%), a validation set (10%), and a test set (10%) to evaluate the training effect and generalization ability of the model.
步骤B3：模型训练；采用Adam优化算法进行梯度下降，初始学习率设定为0.001，Adam结合了动量法和RMSProp的优点，能够在训练过程中自动调整学习率，提升训练速度和稳定性。采用学习率衰减策略，每隔10个epoch将学习率降低一半，以确保模型在接近收敛时能够稳定下降；设置每次训练的批量大小设为64，共训练50轮，每轮包含完整的数据集遍历；采用L2正则化和Dropout技术（比例为0.5）防止过拟合；Step B3: Model training; Adopt Adam optimization algorithm for gradient descent, and set the initial learning rate to 0.001. Adam combines the advantages of momentum method and RMSProp, and can automatically adjust the learning rate during training to improve training speed and stability. Adopt learning rate decay strategy, reduce the learning rate by half every 10 epochs to ensure that the model can steadily decline when it is close to convergence; set the batch size of each training to 64, and train 50 rounds in total, each round contains a complete data set traversal; use L2 regularization and Dropout technology (ratio of 0.5) to prevent overfitting;
步骤B4：模型优化与评估；Step B4: Model optimization and evaluation;
可以使用准确率（Accuracy）、F1-score、ROC曲线等评价指标衡量模型在验证集上的表现。Evaluation indicators such as accuracy, F1-score, and ROC curve can be used to measure the performance of the model on the validation set.
对比实验表明，未经处理的夜视图像在特定条件下的识别率为85%，而经过非对称直方图均衡化处理后，识别率提升至95%。识别速度也得以优化，每张图像的处理时间控制在25ms内。Comparative experiments show that the recognition rate of unprocessed night vision images under certain conditions is 85%, while after asymmetric histogram equalization processing, the recognition rate is increased to 95%. The recognition speed is also optimized, and the processing time of each image is controlled within 25ms.
系统通过OpenCV和Dlib库进行面部检测和特征提取，识别准确率达到了98.5%。这一高精度得益于强大的特征提取能力和模型的深度学习优化。系统包含的报警模块，当检测到儿童进入危险区域（如停车场）或长期停留在异常位置（超过15分钟）时，自动向家长和社区管理人员发送警报信息，短信网关响应时间小于5秒，确保了及时报警。这种快速响应提高了系统的实用性和安全性。系统每月自动选择高质量的儿童面部图像更新数据库中的底图，以适应儿童面部特征的成长变化。The system uses OpenCV and Dlib libraries for facial detection and feature extraction, with a recognition accuracy of 98.5%. This high accuracy is due to the powerful feature extraction capabilities and deep learning optimization of the model. The system includes an alarm module that automatically sends alarm information to parents and community managers when it detects that a child has entered a dangerous area (such as a parking lot) or has stayed in an abnormal location for a long time (more than 15 minutes). The SMS gateway response time is less than 5 seconds, ensuring timely alarm. This rapid response improves the practicality and safety of the system. The system automatically selects high-quality children's facial images every month to update the base map in the database to adapt to the growth and changes of children's facial features.
可以理解，本发明是通过一些实施例进行描述的，本领域技术人员知悉的，在不脱离本发明的精神和范围的情况下，可以对这些特征和实施例进行各种改变或等效替换。另外，在本发明的教导下，可以对这些特征和实施例进行修改以适应具体的情况及材料而不会脱离本发明的精神和范围。因此，本发明不受此处所公开的具体实施例的限制，所有落入本申请的权利要求范围内的实施例都属于本发明所保护的范围内。It is to be understood that the present invention is described by some embodiments, and it is known to those skilled in the art that various changes or equivalent substitutions may be made to these features and embodiments without departing from the spirit and scope of the present invention. In addition, under the teachings of the present invention, these features and embodiments may be modified to adapt to specific circumstances and materials without departing from the spirit and scope of the present invention. Therefore, the present invention is not limited by the specific embodiments disclosed herein, and all embodiments falling within the scope of the claims of this application are within the scope of protection of the present invention.
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202411409750.4ACN118918629B (en) | 2024-10-10 | 2024-10-10 | Community child safety tracking system based on face recognition |
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202411409750.4ACN118918629B (en) | 2024-10-10 | 2024-10-10 | Community child safety tracking system based on face recognition |
| Publication Number | Publication Date |
|---|---|
| CN118918629Atrue CN118918629A (en) | 2024-11-08 |
| CN118918629B CN118918629B (en) | 2024-12-24 |
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202411409750.4AActiveCN118918629B (en) | 2024-10-10 | 2024-10-10 | Community child safety tracking system based on face recognition |
| Country | Link |
|---|---|
| CN (1) | CN118918629B (en) |
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN102376312A (en)* | 2010-07-28 | 2012-03-14 | Lsi公司 | Dibit extraction for estimation of channel parameters |
| CN104299359A (en)* | 2014-10-29 | 2015-01-21 | 合肥指南针电子科技有限责任公司 | System and method for monitoring activity range of children in community |
| CN204374635U (en)* | 2015-01-19 | 2015-06-03 | 成都布林特信息技术有限公司 | A kind of can intelligent children's fever safety monitoring wrist-watch of real-time monitored |
| CN109447048A (en)* | 2018-12-25 | 2019-03-08 | 苏州闪驰数控系统集成有限公司 | A kind of artificial intelligence early warning system |
| CN110491004A (en)* | 2019-08-14 | 2019-11-22 | 金陵科技学院 | A kind of residential communities personnel security management system and method |
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN102376312A (en)* | 2010-07-28 | 2012-03-14 | Lsi公司 | Dibit extraction for estimation of channel parameters |
| CN104299359A (en)* | 2014-10-29 | 2015-01-21 | 合肥指南针电子科技有限责任公司 | System and method for monitoring activity range of children in community |
| CN204374635U (en)* | 2015-01-19 | 2015-06-03 | 成都布林特信息技术有限公司 | A kind of can intelligent children's fever safety monitoring wrist-watch of real-time monitored |
| CN109447048A (en)* | 2018-12-25 | 2019-03-08 | 苏州闪驰数控系统集成有限公司 | A kind of artificial intelligence early warning system |
| CN110491004A (en)* | 2019-08-14 | 2019-11-22 | 金陵科技学院 | A kind of residential communities personnel security management system and method |
| Publication number | Publication date |
|---|---|
| CN118918629B (en) | 2024-12-24 |
| Publication | Publication Date | Title |
|---|---|---|
| CN108446617B (en) | A fast face detection method against profile interference | |
| CN104123544B (en) | Anomaly detection method and system based on video analysis | |
| CN108197546B (en) | Illumination processing method and device in face recognition, computer equipment and storage medium | |
| WO2021208275A1 (en) | Traffic video background modelling method and system | |
| CN104166861B (en) | A kind of pedestrian detection method | |
| CN102903124B (en) | A kind of moving target detecting method | |
| WO2021088300A1 (en) | Rgb-d multi-mode fusion personnel detection method based on asymmetric double-stream network | |
| CN110728223A (en) | Helmet wearing identification method based on deep learning | |
| CN102306304B (en) | Face occluder identification method and device | |
| CN103927520A (en) | Method for detecting human face under backlighting environment | |
| CN108898051A (en) | A kind of face identification method and system based on video flowing | |
| CN114898261B (en) | A sleep quality assessment method and system based on video and physiological data fusion | |
| CN106993188B (en) | A HEVC compression coding method based on multi-face video saliency | |
| CN110390673A (en) | An automatic cigarette detection method based on deep learning in surveillance scenarios | |
| CN112232205B (en) | Mobile terminal CPU real-time multifunctional face detection method | |
| CN114708532A (en) | Monitoring video quality evaluation method, system and storage medium | |
| CN112232204A (en) | Living body detection method based on infrared image | |
| CN110096945A (en) | Indoor Video key frame of video real time extracting method based on machine learning | |
| CN110363114A (en) | A kind of person works' condition detection method, device and terminal device | |
| CN112507818A (en) | Illumination estimation method and system based on near-infrared image | |
| CN108737785B (en) | Indoor automatic detection system that tumbles based on TOF 3D camera | |
| CN110580455A (en) | image recognition-based illegal off-duty detection method and device for personnel | |
| CN116994183A (en) | Multi-algorithm intelligent analysis method and device, storage medium and computer equipment | |
| CN116704579A (en) | Student welcome new photo analysis system and method based on image processing | |
| CN116386275A (en) | Old person's monitoring system that tumbles based on visual identification |
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |