Design And Experimental Research Of Portrait Tracking Device Based On Mobile Observation Platform

Video people tracking has always been a hot spot in the field of computer vision.It's realization faces many challenges,such as random and complex background,occlusion of itself and obstacles,multi-angles and behavioral gestures,etc.,all of which increase the difficulty of the people image extraction and tracking from video.Recently,the novel coronavirus(COVID-19)has caused a huge impact on the health care,financial economy and agribusiness of our country and the world.Wearing a mask is an important and effective anti-epidemic measure.Relevant departments can monitor the wearing of masks in public by video.This is the background of the paper,we study the video people tracking techniques and design an observation and tracking system combining mechanical hardware observation and deep learning algorithm.The mechanical hardware observation equipment adopts a worm gear structure,which can adjust the angle and tilt of the camera to achieve multi-angle framing to obtain stable and high-quality surveillance video.The deep learning algorithm further extracts the portrait from the surveillance video for unique identity recognition,and detects the mask wearing condition of the portrait.Among them,we extract the portrait from the surveillance video,and return the relative position of the portrait in the surveillance video to the hardware equipment,so that it can adjust its angle and pitch to achieve physical-level people tracking.We analyze the performance and influencing factors of the mechanical hardware observation equipment,determined the key parameters of each component,and finally completed the design of the observation system mechanical hardware and put it into use.In the aspect of deep learning video people tracking,this topic designs a portrait feature memory database and a portrait recognition neural network.The portrait feature memory is responsible for storing the recognized portrait features.The portrait recognition neural network is used to complete the matching of the portraits in the current video frame and the portrait features in the database.The video is processed frame by frame and tracks each pedestrian that appears in the video.Further,we consider the secondary tracking matching of forward frames and the tracking optimization technology of forward multi-frames to ensure the accuracy of video tracking while greatly improving the calculation efficiency.In order to realize the detection of mask wearing for tracking portraits,we preprocess the obtained portrait images based on fuzzy geometric segmentation to obtain samples of facial feature classification,thereby constructing a facial feature recognition data set for people wearing masks.Combined with the facial classification model of the convolutional neural network,the goal of accurately identifying whether the portrait is wearing a mask is achieved.This paper combines the mechanical hardware of the observation system,deep learning algorithms,and the image processing strategy in the field of computer vision to achieve efficient and accurate real-time tracking and recognition of specific features of the video.This paper provides a feasible implementation direction for many related application scenarios,and has high practical significance and social value.