The paper analyses the problem of person re-identification accuracy in distributed video surveillance systems by using various datasets for training convolutional neural networks (CNN). After analysis we constructed large joint image dataset of people consisting of CUHK02, CUHK03, Market-1501, DukeMTMC-ReID, MSMT17 and our collected PolReID. PolReID includes 52035 images for 657 people Experimental results on assessing re-identification accuracy based on the main metrics Rank and mAP are presented. The research was carried out for the most widely used CNNs in re-identification, such as ResNet-50, DenseNet121 and PCB. We show that the constructed large dataset allowed us to improve Rank1, mAP for all test sets.