Improvement Of ViBe And Its Application In Person Re-Identification

Moving target detection refers to detecting and extracting moving targets in the video,which is the key to handling video problems.The traditional moving target detection algorithm is difficult to adapt to the dynamic changes in the background,such as illumination,leaves moving with the wind,etc.,resulting in a large amount of noise in the foreground target.The Visual Background Extractor(ViBe)algorithm can quickly and accurately extract the foreground target by using the sample model,but there is a problem that the extracted target is incomplete.Meanwhile,the extracted foreground images,such as pedestrians,involve the matching between images at different viewing angles,which leads to Person Re-identification,that is,the matching research of people's images from different perspectives.According to the question that the extracted target is incomplete of ViBe.In this thesis,sample model is used to construct the background model.Combined with the background difference method,the problem of the incomplete target of the ViBe is improved.After extracting the moving target,the Triplet-ResNet neural network is used to extract and train the extracted pedestrian pictures,and the matching between the pedestrian pictures is completed to identify the same pedestrians from different perspectives.The main research contents include:Firstly,for the ViBe,the moving object was extracted with an "empty hole"or extract an incomplete moving target.The ViBe sample model is used to construct the background model,and the foreground extracted by the background difference method is combined with the foreground extracted by the ViBe to obtain the foreground target.The integrity of the moving target is improvedSecondly,this thesis uses an improved algorithm to extract the pedestrian image into the training set to reduce the impact of over-fitting.Meanwhile,the dropout layer is added to the bottleneck structure of the residual network.The experimental results show that increasing the dropout layer can improve the accuracy of the residual network in Person Re-identificationThirdly,the Triplet-ResNet network is used for pedestrian recognition.The residual network of the dropout layer is added as the skeleton network,and the SoftMax loss and Triplet loss training network model are combined to effectively extract pedestrian characteristics.The experimental results show that the Triplet-ResNet network is used.The accuracy of the traditional algorithm for pedestrian recognition is improved.