Pedestrian Monitoring Based On Low Altitude Dwell Platform And Deep Learning

With the rapid development of the city,there will be many urban traffic problems.It is necessary to monitor and obtain urban traffic information in time and effectively.Among them,remote sensing technology has been paid more and more attention in monitoring urban traffic information.Pedestrian information is the basic information in urban traffic system and pedestrian monitoring is an important research content in intelligent traffic system.This paper proposes a new and convenient way of urban remote sensing monitoring,which is low altitude dwell platform.This monitoring method has the advantages of large coverage,low acquisition cost,free choice of high and low level,and real-time observation.However,the traffic pedestrian video extracted from low altitude dwell platform also has the characteristics of complex scene,different target sizes,and difficult traditional recognition methods.In recent years,with the emergence of a variety of data sets and the continuous improvement of computer performance in the era of big data,deep learning has made remarkable achievements in computer vision.Convolutional neural network has the characteristics of strong generalization ability and high robustness in target detection.Therefore,this paper chooses low altitude pedestrian video combined with deep neural network for pedestrian monitoring.In the research of low altitude pedestrian video monitoring,this paper focuses on two situations: sparse pedestrian and dense pedestrian.Through reading literature and experimental processing,several methods for pedestrian detection under low altitude dwell platform are summarized.Three main algorithms,namely yolov3,Faster R-CNN and SSD are used for sparse population combined with migration learning,and two kinds of neural network structure model algorithms,namely encoding-decoding and convolution+cavity convolution are used for dense population.In the process of the experiment,we first annotate the pedestrian data set collected by self collection and extensive collection,and then complete the experiment by image preprocessing,model building,training model,test model to complete the experiment.The experimental results show that the accurate detection of pedestrian target is achieved.The recognition accuracy of the model in sparse pedestrian detection reaches 94.5%,and the MES and MAE index in dense crowd counting reaches 114 and 66.7.The model trained from the experimental results has high accuracy,good robustness and wide range of use.Pedestrian detection model is the premise of counting and tracking.The pedestrian counting method used in this paper is based on the virtual detection line,which can realize the two-way counting through the detection line.The calculation method of pedestrian flow is based on the target tracking algorithm based on Kalman filter,and the average detection accuracy of four videos is 91%.After the model training and testing,take photos of pedestrians,road traffic pedestrians and dense crowd activities on campus for scenario application analysis.Explore the feasibility of monitoring urban pedestrians on low-altitude parking platforms,provide reference value for monitoring cities using remote sensing and verify whether deep neural network technology meets the needs of low-altitude pedestrian video monitoring.