Research And Implementation Of Elevator Passenger Abnormality Detection Method Based On Edge Intelligence

With the increasing speed of modernization in my country,the market demand for elevators has increased year by year.At present,the mainstream elevator passenger abnormality detection system is based on artificial intelligence technology based on network communication and server architecture.The edge intelligence technology that appeared in recent years does not require network communication,has high reliability,and is low in cost.Therefore,this thesis proposes a method for detecting anomalies in elevators based on edge intelligence.Usually,the deep neural network has a large amount of computation and needs to be deployed on high-performance servers.The edge-intelligence-based elevator passenger anomaly detection method proposed in this thesis can be directly deployed on the edge intelligent devices.The method is modified from YOLO v3 by optimizing inference speed based on Roofline model of deployment equipment,named as YOLO v3-R.The optimization procedure is shown as: Firstly,selecting the edge intelligence device and establishing its Roofline model;then,choosing a variety of network convolution structures,performing their classification layer by layer according to the work area of Roofline model,and using different parameters to evaluate their inference time;finally,picking out the network structure with the shortest inference time for network reconstruction.The Pascal VOC2007+2012 dataset is used to train and test the YOLO v3-R network.The results show that the average detection accuracy of the YOLO v3-R network is comparable to the original YOLO v3,while the image inference time is reduced by 26%,and the number of real-time video inference frames is increased by24%.For the scenario of elevators,a special dataset APCE is created for the training of the passenger anomaly detection network YOLO v3-R.Meanwhile,an experimental system is designed and assembled for detecting the abnormality of passengers on elevators.The Jetson Nano is used as the hardware platform of the system,and YOLO v3-R is adopt as its target detection algorithm.The detection process is as follows: using the ray method to detect the effective range of the prediction frame output by YOLO v3-R and recording the passenger status flag bit;according to the recorded results,determining whether there is an abnormal passenger loading event.The experimental system shows that the real-time reasoning speed can reach 1.96 frames per second without network communication,and reaches the expected target.