Abstract
Person re-identification from video is very challenging due to the problems arising from different viewpoints, illumination variations, background subtraction errors or human detection etc. In order to improve the performance of person re-identification in surveillance video, we propose to incorporate the human height information and temporal consistency to enhance the human detection in the person re-identification system. The proposed system consists of the following steps: First, we employ background subtraction to obtain the foreground regions and apply the human detector to some beginning frames to estimate the parameters for human height projection onto images. Thus, we can establish the relationship between image positions and the sizes of sliding windows for human detection. Second, we apply the sliding window search on the foreground region by computing the associated color and texture histograms. Thus, the probability of human detection with the SLBP features is computed to improve the human detection accuracy. Finally, we compare the features and compute the similarity score with adjustment based on temporal consistency. Experimental results show superior performance on person re-identification by applying the proposed algorithm that utilizes human height information and temporal consistency to some publicly available datasets.