Research On Real-Time Detection Method Of Face Occlusion Based On SSD And System Implementation

With the rapid development of artificial intelligence,intelligent detection systems with identification and processing functions have received more and more attention.The abnormal behavior detection of face occlusion is one of the important research topics.In daily life,criminals usually covering their facial features to conduct crimes.Therefore,abnormal behavior detection of face occlusion can effectively assist in the prejudgment of criminal behavior.However,the existing abnormality detection of face occlusion mainly has the following shortcomings:(1)Most existing methods use the method of human abnormal behavior detection with multi-step detection,Therefore,this kind of method show poor robustness in the complex environment of multiple occlusion targets.(2)Existing abnormal behavior detection algorithm based on traditional neural network has the problems of high computational complexity,numerous model parameters and poor real-time performance.(3)Existing deep learning-based anomaly behavior detection methods only stay based on the desktopside demonstration of the deep learning open source framework,due to the inability to break away from the open source deep learning framework,makes such algorithms poorly usable and poorly portable.In order to solve the problem that the existing face occlusion abnormal behavior detection method exhibits poor robustness in the complex environment of multiple occlusion targets,this paper proposes a face occlusion abnormal behavior detection algorithm based on the target detection framework SSD.The algorithm firstly detects the human face coverings(including masks,sunglasses,etc.)through the multi-target detection framework,so as to achieve the purpose of extracting facial features of the human body.Then,the extracted facial features of the human body are used to determine the occurrence of abnormal behavior of the face occlusion.The implementation shows that the proposed method for detecting abnormal occlusion behavior can detect the occurrence of multi-person anomalous behavior in complex environments and achieve high accuracy on self-recorded data sets.In order to solve the problem of the existing human face occlusion behavior abnormal behavior detection method,the computational complexity is high,the model parameters are large,and the real-time performance is poor.Compare and analyze network models such as VGG,MobileNet,and GoogleNet under different computing power and bandwidth.This paper proposes a face occlusion abnormal behavior detection algorithm based on lightweight SSD.The algorithm discards the target detection framework SSD based on the traditional convolutional neural network VGG,and uses a lightweight MobileNet network to extract features.Experiments have shown that the speed of the SSD used in the CPU is 8.5 times faster than that of the VGGbased SSD.In order to solve the problem of poor practicality and poor portability of existing face occlusion abnormal behavior detection methods,this paper designs and implements a face occlusion abnormal behavior detection system,which improves face occlusion detection through face occlusion determination module.The accuracy of the system is up to 95%,and it is out of the Caffe framework.It can detect the abnormal behavior of face occlusion in real time in the CPU,which has strong practicability and flexibility.