Person re-identification from multiple surveillance cameras combining face and body feature matching

Abstract

In this work, a new method for human re-identification (Re-ID) from multiple surveillance cameras is proposed. Unlike traditional methods in which only the body features are used for matching, our proposed method uses both body and facial features for the Re-ID process. This combination allows us to re-identify people in challenging conditions, such as people with a uniform or similar-looking outfit, a partial or occluded body, appearance changes, or illumination. The face and body feature extraction models were developed using the state-of-the-art deep neural backbones and the synthesis of existing datasets in the literature. The performance of the method was evaluated on a self-generated dataset, which contains images under challenging conditions. The evaluation results show that our method outperforms traditional methods, in which the accuracy Rank1 reaches 91.30% while the traditional ones have a Rank1 of only 86.96%. This newly introduced method can be used for many practical applications in security surveillance of buildings and offices where challenging conditions often appear.

An efficient pedestrian attributes recognition system under challenging conditions

“…For practical applications in security surveillance scenarios, 13 typical attributes of the human body outfits are chosen to label, including "Male", "Female", "Adult", "Children", "Long-Hair", "ShortHair", "Hat", "Long-Sleeves", "Short-Sleeves", "Trouser-Jeans", "Skirt", "Short", and "Backpack". Since images of PA100k, MSMT17, and Market1510 are not assigned to person IDs, the re-identification method from [26] was used for the person ID register. The EPAR dataset contains 90 004 images of 7445 person IDs, with annotations for 13 binary attributes.…”

“…Also, this model outperforms others in accuracy and robustness in challenging detection scenarios. Since the images of a detected person can appear in several video frames of different cameras, the re-identification method [26] is used to match these images together and assign them to a tracking ID. Consequently, images and poses of each tracking ID with different viewpoints from different cameras are continued processing in the following steps.…”