基于人脸与步态特征的室外作业场景身份核验方法

Identity verification method based on face and gait features in outdoor operation scenes

在建筑、通信、电力等工程行业中,作业人员需要频繁执行室外作业。由于室外环境复杂,许多运维工作存在来自高压、高空、深坑等因素的高风险。安全事故一旦发生,将造成巨大的人员和财产损失。因此,需要在作业过程中对作业人员进行身份核验,以方便监督。然而,在传统的监督方式中,作业现场的人员管理和行为管控完全依靠人工核查,监控视频也依赖人工看守,无法做到人员身份实时核验以及对非作业人员入场的有效预警。针对室外作业场景中的作业人员身份识别,目前的研究方法大多基于人脸识别。人脸识别方法能够在作业人员脸部信息清晰且完整时准确识别出其身份信息。然而,当存在遮挡,以及受检测距离、检测角度等因素影响时,会造成脸部信息不完全或者模糊,导致采用人脸识别方法难以准确识别出作业人员的身份。步态特征是一种描述行走方式的复杂行为特征,包括脚的触地时间、离地时间和人体高度、双手摆动幅值等。相比人脸识别,步态特征识别有以下优点:第一,步态识别适用的检测距离更远,而人脸特征随着检测距离的增加识别难度明显上升;第二,步态特征识别是非主动识别,现场作业人员几乎随时随地处于行走状态,而人脸识别需要识别对象正对检测装置;第三,步态特征具有较强的特异性,不像人脸特征较易被模仿、修改。不过,单独采用步态特征进行识别,虽然在作业人员运动时可以捕捉体态信息进而较为准确地对其进行身份核验,但是无法对静态的作业人员进行身份核验。针对此问题,提出了一种多特征融合的身份核验方法,结合步态与人脸等多特征进行识别,不受衣着、环境等外在因素的干扰,可以有效提高身份核验的准确率。提出的融合人脸特征和步态特征的多特征身份核验方法包括身份注册阶段、训练阶段与测试验证阶段。身份注册阶段,人工标注人脸与步态信息,并录入人员信息库;训练阶段,首先利用相关网络提取视频中图像序列的步态轮廓图与人脸区域,然后利用深度学习网络模型提取相关特征,构建融合的特征向量与身份ID间的关系;测试阶段,首先判断图像中有无清晰人脸,如果有则使用多特征融合识别方法,否则仅通过步态特征进行特征匹配进而完成身份核验。结果表明,多特征融合方法在中科院自动化所的CASIA-A数据集上的分类准确率达到99.17%,数据集包含的3个视角下的分类准确率分别为98.75%,100%和98.75%。因此,所提方法可以有效提高单人场景中的身份识别准确率,是在室外作业场景中进行身份核验的一种有效方法。

In construction, communications, power and other engineering industries, workers need to perform outdoor operations frequently.Due to the complex outdoor environment, there are high risks from factors like high voltage, high altitude, and deep pits in many operation and maintenance tasks.Once an accident happens, huge losses in personnel and property would be caused.Therefore, it is necessary to verify identities of operators during the operation process for supervision.However, in the traditional supervision method, the personnel management and behavior control in the operation scenes rely entirely on manual verification, and the surveillance video also relies on manual guards.It is impossible to achieve real-time verification of personnel identities and effective warning of the entry of non operators.For the identification of workers in outdoor work scenes, most of the current research methods are based on the face recognition.Face recognition method can accurately identify the identity information of a worker when his facial information is clear and complete.However, when the facial information is incomplete or fuzzy because of occlusion, long detection distance or the inclined detection angle, it will be difficult to accurately identify the operator’s identity with the face recognition method.Gait feature is a complex behavioral feature of a walking person, including the time the foot touch and leave the ground, the human height, and the swing amplitude of hands.Compared with face recognition, gait recognition has the following advantages.Firstly, the distance applicable to gait recognition is longer, while the recognition difficulty of facial features increases as the detection distance increases.Secondly, gait feature recognition is non-active, and workers on the operation scene are walking almost anytime and anywhere, but face recognition expects the recognition object to face the detection device.Thirdly, gait features have strong specificity and are difficult to be imitated and modified.Nevertheless, gait information cannot be used alone for identity verification of workers in static poses.To solve the above problems, a multi-feature fusion identity verification method was proposed, which combined multiple features such as gait and face features for recognition without being interfered by external factors such as clothing and environment.This method effectively improved the accuracy of identity verification.This multi-feature identity verification method, combining face recognition and gait recognition, included identity registration phase, training phase and test phase.In the registration phase, the face and gait information were manually marked and recorded in the database.In the training phase, the correlated network was used to extract the gait contour map and face region of the image sequence in the video.Then the deep learning network model was used to extract relevant features in order to build the relationship between the fused feature vector and the identity ID.In the test phase, whether there is a clear face in the image was judged.If so, the multi-feature fusion recognition method was used.Otherwise, only the gait feature for feature matching was used to complete the identity verification.The results show that the proposed multi-feature fusion method achieves the classification accuracy of 99.17% on the CASIA-A data set of the Institute of Automation, Chinese Academy of Sciences.The classification accuracy is 98.75%,100% and 98.75% in the three views included in the dataset.Therefore, the proposed method can effectively improve the accuracy of identification in single-person scenes, thus providing an effective scheme for identity verification in outdoor work scenes.