Algoritm Research Of Pedestrian Detection And Tracking Under Occlusion Scene

Nowadays,the statistics of information data have become an important task which plays an important role in many aspects.Among them,the statistics of the flow of people in the commercial sales model to adjust and the safety of public utilities plays an important role,such as commercial competition,so the use of traditional manual experience of passenger flow statistics and take it as reference is not accurate enough.In many other public places such as railway stations and scenic spots are often prone to be crowded during rush hour,the lack of monitoring of passenger flow in these environments also makes more and more risk.Therefore,the statistics of passenger flow in public places has become a real demand.The key to complete passenger flow statistics is detecting and tracking pedestrians accurately,of which pedestrian detection is the basis of pedestrian tracking.The current pedestrian detection and tracking methods face the following difficulties:(1)The final detection score of the DPM model used by researchers previously is a combination of the root filter and the component filter score.For the occluded target image,the calculation of the target characteristics will be affected by the occlusion part of the characteristics will be obscured and impacts filter in corresponding position response score,resulting in deviation of location of target or missing of target;(2)Training time of target detection model based on deep learning is too long under the condition of non-mini batch.Otherwise the detection accuracy of the model improves in most cases by batch normalization,but basically fails when dealing with mini batches of samples and non-independent and identically distributed data;(3)Particle filter algorithm also has some drawbacks when the target is blocked heavily.For example,the lack of diversity of particle samples often occurs,and the surplus statistical calculation of particles with higher weights leads to the failure of the target tracking and deviation of location.The contribution and innovation of this thesis include the following three aspects:(1)Aiming at the problem that pedestrian occlusion leads to the low score of component filter,this thesis propose a weighting component model for DPM model.After the DPM training phase is completed,the weight of each component filter is tested and calculated,then the weighted component model is used to detect pedestrians which are partly occluded and get good result;(2)Aiming at the problem of low detection accuracy of YOLO v2 model,this thesis introduces the batch re-normalization process to improve the structure of YOLO v2 model,combining the advantages of batch normalization algorithm to deal with small batch samples and non-independent and identically distributed data.That is,regard the feature map produced by the convolution layer during convolution operation as a neuron and normalize it,which effectively reduce time cost of training and improve the detection accuracy of the model;(3)Aiming at the poor tracking performance of the particle filter when the pedestrian is obscured,BP neural network algorithm and typical sampling algorithm are combined in this thesis to increase the weight of the particle located at the tail of the probability distribution to enter the high-weight area so as to improve the filtering accuracy.