Indoor Pedestrian Detection Based On Nuclear Density And Neural Network

Pedestrian detection is an important part of intelligent assistant driving,intelligent monitoring,pedestrian analysis and intelligent robotics.Pedestrian testing has entered a stage of rapid development since 2005,but there are still many problems to be solved,such as the unsupervised situation in certain indoor specific places(merchants,shopping malls,warehouses and outbound tourist families).The situation has major security risks,and these hidden dangers mainly come from people.If the manual supervision of these places will cost a lot of manpower and material resources,and can not be monitored from time to time,it will bring opportunities for criminals,causing losses to businesses and individuals.This thesis mainly proposes an indoor pedestrian detection method for this problem,which can effectively monitor these places automatically without being taken care of.This thesis focuses on the improvement of the speed and false alarm rate of the pedestrian detection algorithm.The main research contents are as follows.First,in the detection process of pedestrians,the moving target detection method based on nuclear density will generate pixel point misidentification problems,which will affect the detection results.This thesis proposes to improve the kernel density method by applying the inter-frame difference method to solve the problem of pixel point misidentification.The effectiveness of this method is verified by relevant experiments.Secondly,the motion detection algorithm can only detect and track moving targets,and cannot identify moving targets.While neural network methods such as Yolov3,Tiny-Yolov3 and Retinanet can identify pedestrians,it is difficult to ensure that the detection speed and accuracy are not met in complex scenes,and no false alarm targets are present.The Yolov3 algorithm has a high accuracy rate,but it also has limitations.When using the Yolov3 method for pedestrian detection,the image frame is detected regardless of whether the image frame contains a target,which reduces the efficiency of pedestrian detection.Aiming at these problems,this thesis proposes an improved combination of kernel density and Yolov3,speeds up the efficiency of the Yolov3 algorithm for detecting pedestrians,and uses a combination algorithm to remove a large number of false targets and reduce the false alarm rate of the algorithm.Finally,20 sets of videos were selected from the dataset2014 database,network and self-built database for experiments.The experimental results show that the algorithm improves the detection speed by 2.94 times and the false alarm evenly when the detection accuracy is constant.The rate of false alarm rate of Yolov3 algorithm has dropped significantly,up to 37.05 times,the false alarm rate dropped to 0.105%,the minimum dropped by 9.625 times,and the false alarm rate dropped to 0.64%.