Study On Particle Filter Object Tracking Algorithm Based On Convolution Neural Network Feature

In recent years,the intelligent video surveillance technology based on the artificial intelligence has been developed rapidly.Traditional tracking algorithms usually use artificial characteristic,which based on the prior knowledge,to establish the target feature models.Because of ignoring the essential feature of target,the adaptability of traditional tracking algorithm is not robust.In order to obtain the robust feature model,the convolutional neural network is used to learn high-level semantic features of targets instead of traditional lowlevel features.Two kinds of appearance models are designed based on convolutional neural network,classification network model and contrast network model.Correspondingly,two target tracking algorithms are established under particle filter tracking framework in this thesis.The main achievements are presented as follows:First of all,for the robust appearance model of tracking target,a small classification model is designed based on convolutional network,which gets the deep features from moving target.Aiming to solve the problem of few tracking target samples,the proposed classification network model combines the off-line training and on-line fine-tuning method to adapt different tracking targets.In addition,because the model updating of classification network is time-consuming,a contrast network model is presented based on siamese convolutional networks.This model is trained on off-line using video sequences samples,which achieve a similarity measure function to adapt the tracking problem.The contrast network model improves the computational efficiency because it needn't to be updated online.Secondly,combining with the particle filter tracking framework and the deep network feature models,two target tracking algorithms are established.The tracking classifier is designed to track the moving object with the classification network model.In this tracking algorithms,model update strategies which include long-time and short-time online updating procedure are adopted to enhance the anti-interference ability under target posture change,similar background,and so on.The online hard example mining strategy also used in this algorithm to improve the online learning efficiency.Furthermore,the tracking algorithm based on contrast network model uses the similarity function to perform the measure between the target and candidates.A tracking template updating strategy is established to solve the target occlusion and other problems.Finally,the proposed tracking algorithms are performed using Caffe deep learning framework and Matlab2014 a under the Ubuntu 14.04.The simulation results show that the proposed algorithms based on convolutional neural network can realize robust tracking in some complex circumstances,such as occlusion,illumination changes,pose variations.