Non-contextual Long-Term Person Re-ID Under Sudden Illumination Conditions

Abstract

AbstractPerson Re-ID (Re-ID) has been an emerging topic and there is a need for non-contextual, long term person Re-ID algorithm since most of the crimes occur in public places such as airports, railway stations where the video is recorded for a long duration across arbitrary camera views. The proposed work comprises of two fusion frameworks. First, the people detection is accomplished by fusing Histogram of Oriented Gradients (HOG) and extended Center Symmetric Local Binary Pattern (XCS-LBP) features which overcomes several disadvantages like missing detection and false detection due to sudden illumination, and near/far field of view changes. Next, each body part of the tracked person is learned using Deformable Part Model (DPM) which is robust to different viewpoints. Secondly, the feature level fusion of Gabor (appearance feature) and Skeleton Recurrent Motion Image (SRMI Gait biometric feature) is proposed to overcome the homogeneity issue of distinguishing people when both the texture and color of the people attire are similar. Finally, person Re-ID is achieved by relevance metric learning method with list wise constraints (RMLLCs). The performance measure, Cumulative Matching Curve (CMC) Rate shows the improved matching accuracy compared to other state-of-the-art algorithms with benchmark datasets.KeywordsPerson Re-identificationLong-termSkeleton recurrent motion image sudden illuminationGait and appearance features