Pedestrian Safety Detection System On Escalator Based On Human Pose Estimation

With the rapid development of China’s economy,escalators have gradually become an indispensable part of people’s lives in recent years.However,escalators may do continuous and serious harm to people on the condition that pedestrians have an accident such as a fall,for the reason that escalators run without interruption for a long time,which relies on manual video surveillance to observe possible dangers as early as possible.This method causes problems such as high labor cost,slow response speed,and many false or missed detections.Therefore,it is of great social significance and application value to do research on the escalator pedestrian safety detection system with high accuracy,strong real-time performance,and no manual labor.This thesis studies the related algorithms of pose estimation,target detection and classification,and designs an escalator pedestrian safety detection system.The works done in this paper are as follows:(1)Study and compare the principles and performances of various human pose estimation algorithms and choose OpenPose as the pose detection algorithm for the reason that OpenPose not only has high accuracy,but the detection speed is also beyond the reach of other algorithms.In view of the complex problem of predicting keypoints in OpenPose,a set of post-processing procedures is proposed,including discarding keypoints according to confidence,reorganizing keypoints,eliminating data with too few keypoints,and normalizing keypoints.The cleaning work has been carried out to improve the scientificity and validity of keypoints.In the test environment,the m AP of the OpenPose algorithm reaches 88.6%,and the processing speed is151.1ms/frame.(2)In order to reduce the false positive of OpenPose,propose a way to combine the target detection algorithm with the pose estimation algorithm.By comparing the performance of various target detection algorithms,the YOLOv4 algorithm is chosen as the target detection algorithm for its accuracy and speed.Accuracy of YOLOv4 is improved by expanding the dataset and regenerating anchors.Finally,the m AP of the YOLOv4 algorithm reaches 82.5%,and the false positive rate of OpenPose algorithm is reduced to 0 from 2.5% in the test environment.(3)Study and compare a variety of classification algorithms,and construct the dataset by combining public and self-made datasets.In the case of dataset testing,the accuracy of the MLP classifier reaches 90.73%,higher than 74.69% of the SVM.Therefore,the MLP classifier is selected as the classification scheme of the algorithm to realize the classification of human pose.(4)Design an escalator pedestrian safety detection system based on the C/S architecture.The server uses the combination of OpenPose and YOLOv4 algorithm to achieve high-precision detection of keypoints of people,and then uses the MLP classifier to classify the keypoints of people detected on the escalator.Once there is a dangerous pose,an alarm will be given and the corresponding images and videos will be saved.The client is responsible for receiving the alarm information from the server and checking the alarm information accordingly.According to test results,under the GTX 1060 graphics,the system can process 1920×1080 images at 4 frames/s,and the accuracy can reach 94.7%.The system meets the requirements of the monitoring system for speed and accuracy.