Research On Key Technology Of Person Re-Identification

With growing concern for public safety and development of video capture and large-scale data storage technology, a large amount of surveillance cameras are set in the places where public safety incidents could easily take place, such like shopping malls, parks, schools, hospitals, companies, sports stadiums, plazas and subway stations. With the amount of video boosting, watching the surveillance video becomes a tedious and troublesome work. A need of re-identifying person in surveillance video by using computer vision technology emerged, which is termed as person re-identification problem. More specifically, person re-identification is to identify whether a detected person has emerged before in the surveillance network. In this thesis, two re-identification approaches based on appearance of person are proposed. The main works and contributions are:1. Design a person re-identification algorithm based on color and texture feature. In this algorithm, we modulate the HSV color model, which, to some extent, eliminates the influence on the visual representation caused by illumination variations. To speed up the re-identifying procedure, a fast foreground locating method based on color projection and body-part location constraint is adopted, which reduced the amount of foreground candidates. A reasonable tradeoff is made between re-identification accuracy and algorithm complexity.2. By segmenting the appearance of person into several visual patches using super pixel algorithm, a local feature based on super pixel segmentation is proposed. This feature is robust to key point location-drift which is commonly produced by other local features when the poses of persons change. The person re-identification problem is transferred into image searching problem using method similar to text retrieval by adopting TF-IDF schemes and vocabulary tree method. The speed, as well as the accuracy, of the re-identification is improved. Experiments of algorithm on ETHZ dataset have shown that our method demonstrates a reasonable performance, which can be inferred from the CMC curve indicating average90%re-identifying accuracy in top10ranks. Our method outperforms the state-of-art identification methods, SVM, PCA and PLS algorithm.