Student: Jefferson Colares de Paula
Title: Long Term Person Re-identification Using Multimodal Features
Advisors: Diego Barreto Haddad (advisor), Douglas Oliveira Cardoso (co-advisor)
Committee: Diego Barreto Haddad (president), Douglas Oliveira Cardoso (CEFET/RJ), Fernanda Duarte Vilela Reis de Oliveira (UFRJ), Eduardo Bezerra da Silva (CEFET/RJ), Gabriel Matos Araujo (CEFET/RJ)
Day/Time: December 30, 2020 / 14h.
Abstract: Person re-identification (ReID) consists in comparing images containing people, acquired by multiple cameras with non-overlapping views, and inferring if the people in the images are the same. The problem is more complex than it seems, because the images are prone to great differences in illumination, viewpoint, camera optics, and also partial occlusion, self-occlusion, confusing backgrounds and other challenges. Long-term re-identification, which is the subject of this work, is characterized by the occurrence of an interval between the image captures. This interval has no specific duration, but, in general, it’s greater than 24 hs. In such period, the person being reidentified might have changed clothes or have some minor changes in appearance. This long interval adds complexity to the original ReID challenge because the colors and textures of the clothes, which are the most used features for person reidentification, can’t be used as discriminative elements. This work investigates a solution to the problem of long-term person re-identification by using gait and face characteristics of the person as input features for a machine learning model based on neural networks. The hypothesis to be evaluated is that the combination of gait and face characteristics can have the machine learning model to ignore or minimize the effect of changing clothes and, at the same time, value the motion characteristics. Our results show that the combination of multi modal input characteristics improves the performance of short-term person reidentification and can also be used for long-term person reidentification.