Research And Application Of Object Detection In Elevator Based On Deep Learning

Our daily lives have been changed constantly due to the improvement of emerging technologies in the past decades.The era called"big data era" is coming now.How to collect information fast and correct is very important for the application of "big data" technologies.For example,image data which acquired by the intelligent video camera is primary information in public monitoring.An intelligence monitoring and detection system based on deeping learning method for collecting image informations of passengers in a public elevators was developed in this dissertation,which can realized gender and age classification and statistics.The above system have potential application in many field such as market research,security monitoring.The detection system proposed in the dissertation consists of target detection module and target tracking module,the detection module can obtain facial information quickly and accurately.In this paper,analysis was made towards various detecting algorithms appearing in the deep learning field,considering instantaneity and accuracy of detection algorithm,it was decided to make improvement based on YOLOv3 algorithm and by combining the elevator internal environment,the improvement only contained two aspects as required,the first was to substitute IOU loss function that is common in YOLOv3 algorithm into GIOU loss function,at the same time,improve the bounding box generation in YOLOv3 algorithm by combining Gaussian YOLOv3.The algorithm after the two improvements are called as YOLOv3-B,compared to YOLOv3 initial algorithm,its test accuracy increased by 11%for the self-built data set,while its verification accuracy increased by 4.9%for the classic data set of target detecting algorithm,and the detection was of high robustness under the circumstance that the face is shielded.On the basis of algorithm improvement,in order to further fit the elevator internal environment,pictures inside the elevator were cut and marked on the basis of existing monitor videos inside the elevator,meanwhile,to enrich the data set for training,StyleGAN2 algorithm was utilized to make corresponding face database to improve the training quality.After that,for the sake of preventing repeated counting,face tracing and counting were achieved in this paper by combining DeepSORT algorithm,and the system was stable during counting and could meet the passenger flow information statistics in the elevator,thus to provide reference for passenger information research in the elevator.After the test,this system could meet the work requirements required for enterprise market survey in detection instantaneity and accuracy.