Research And Implementation Of Perspective Invariance Crowd Counting

Population counting is also called crowd counting as a process of using images or video itself to make statistics or estimation of population attributes in images or video.It is a branch of population detection and is closely related to statistics,pattern recognition,machine learning,signal processing,computer vision and other disciplines.In recent years,with the rise of machine learning and deep learning algorithm,crowd counting have gradually become the current hot research direction.This paper mainly elaborates three kinds of improved crowd counting algorithms,namely: 1.People counting based on regression;2.People counting algorithm based on human detection network;3.Population density estimation algorithm based on full convolution network.The aim of the improvement is to ensure the real-time performance of the algorithm,to improve the algorithm's response to the scene and to detect the deformation of the population because of the perspective and perspective.The specific work is as follows:1.Improvement of person counting algorithm based on regression method.Improvement of hybrid gaussian background modeling: the spatial grayscale information is added on the basis of traditional pixel sequence gray information,making the foreground image extracted by the background modeling algorithm more accurate.Improvement in HOG and LBP algorithm of feature extraction algorithm: introduction of representative scenarios perspective perspective density figure,for each pixel position with a weight of perspective,finally calculated characteristics also have a certain perspective invariants.Build a regression network based on one dimensional convolution,and use the causal convolution method in GuGe wavenet to process the feature vector,making the model accurate.2.Improvement of crowd counting algorithm based on human detector.Improve the RCNN candidate area of the detection model generation algorithm: the candidate area background segmentation based on KNN to the prospect of a marker for sliding traversal,and will be used in calculating the candidate area size linear model fitting scene in an offline training perspective,to obtain the candidate area can contain more accurate human body areas.Simplify the adjustment of Alex Net: the purpose of simplifying the network is to make the previously classified network adapt to only the human detection application,and improve the detection efficiency and accuracy.3.Improvement of estimation algorithm based on crowd density.Human density marker algorithm based on gaussian mixture model: tag samples by density of the human body,make samples have more abundant information,such as contain human body area information,population information.Csonstruct FCN-8 network: after the image input network,it will experience 3 pooling operations,and then 3 times of deconvolution operation,reducing the original size to calculate the secret plot.