Research On Algorithms Of Person Re-Identification Based On Surveillance Videos

The video surveillance system is an important security aiding application in urban public areas.In researchs of monitoring,pedestrian tracking and retrieval is the primary issue,and the person re-identification problem is the vital part.Person re-identification technology is the process of extracting features from a series image frames which cut out from surveillance videos for a specific pedestrian target and find out the same targets in the other surveillance videos.The traditional ideas of person re-identification are to find some differentiated and stable features,such as color,texture,image blocks and symmetries.Then,figure out a metric learning function to make the smaller distance between the same pedestrian and the larger distance between different pedestrians.The traditional methods extract dimension feature manually,which is more specific and accurate for the single dimension.However,it is a little bit limited in high dimension feature extractions.Besides the fusions are tough,and the comprehensive features performance is insufficient.In recent years,deep learning,which is popular in computer vision,has also been applied to the tasks of person re-identification.Both pedestrian feature extractions and similarity matching work have been greatly improved compared with traditional methods.After analyzing the currently popular siamese models,lots of irrational parts had been uncovered.This paper divided the original network into two sub modules: the motion network and the spatial network.In the motion module.This paper takes advantage of the Epic Flow which performs better than Optical Flow to improve the accuracy in extracting motion features,and reuses the Epic Flow to filter pedestrian frames for key images by removing the poor performance of the image frames.The model shortens the network computing time with a small amount of key frames to extract pedestrian spatial features.Then,the improved model is used for training.Any pair of pedestrians are chosen to extract the features of two the pedestrian frames sequences and calculate the distance between the two features simultaneously.At last achieve a great classification model with the loss function of the distance between two feature vectors.This research mainly takes the PRID2011 dataset and the i-LID-VID dataset for experiments.The experimental results verify that the improved two-stream siamese model can effectively enhance the accuracy of pedestrian re-identification matching problems.The Rank-1 accuracies on the two datasets were 85.00% and 66.67%,respectively,which were 15.00% and 8.67% higher than the original network before the improvements.