A Study Of Computer Vision-based Escalator Fall Detection Algorithm

With the improvement of people’s living standards and the increasing government investment in infrastructure construction,escalators have become an important and indispensable part of large buildings.Subway stations,airports,large shopping malls and other places with large flow of people are equipped with a large number of escalators.Escalators have a large transportation capacity and can transport passengers continuously,which brings great convenience to people’s life.However,because the escalator is open,the probability of accidents is much higher than that of vertical elevators.Because of the sudden nature of escalator accidents,a fast,accurate,and real-time detection method is urgently needed.However,the traditional escalator detection technology basically detects the mechanical operation status of escalators,such as step deformation,skirt deformation,etc.,but rarely detects the dangerous riding posture under the normal operation status of escalators,such as human falls.Therefore,this paper focuses on computer vision based on human’s dangerous riding posture detection when riding escalators.The main research works of this paper are as follows:(1)Since there is no publicly available dataset for fall detection in escalator scenarios,this paper acquires video images of passenger postures in real escalators and builds our own escalator fall detection dataset through the following processing steps: first,data preprocessing is performed on the images captured by surveillance cameras,including frame extraction and manual annotation.Second,different installation angles of surveillance cameras are simulated.Rotation and mirroring are used for data augmentations.Finally,for the noise generated during image acquisition,Gaussian filtering and median filtering are used,and bilinear interpolation algorithm is used to maintain the resolution while compressing the images to obtain better image quality of the dataset.(2)A fall detection model based on an improved Single Shot Multi Box Detector(SSD)algorithm is proposed to address the problem of accurately identifying passengers’ dangerous postures in real escalator scenarios.First,the model adopts the lightweight Mobile Net V2 as the backbone network,which has a simpler structure and faster detection speed.Second,the attention module is added to the model,which further enhances the semantic information of the high-level feature maps without significantly increasing the computational cost.Finally,the loss function is optimized.Experimental results show that the improved SSD fall detection algorithm is able to identify the dangerous posture of escalator passengers accurately and in real time.(3)A fall detection model based on the improved YOLOv5(You Only Look Once version 5)algorithm is proposed to address the problem of different scales of detected targets caused by the installation distance of surveillance camera.First,the network structure is improved according to the requirements of real-time and accuracy of fall detection,and optimized by constructing a lightweight network structure.Second,the loss function is improved by considering the regression aspect ratio of the bounding box.Finally,the adaptive feature fusion network is introduced to enhance the model’s ability to fuse multi-scale features,considering the problem of different scales of targets existing in real scenes.The experimental results show that the improved YOLOv5 fall detection algorithm improves the detection speed while maintaining the detection accuracy compared with similar single-stage target detection algorithms.