An Efficient Gmm And Active Contour Based Unsupervised Person Re-Identification

This paper proposes a novel algorithm for static and single camera foreground detection and multi-person tracking using active contour and Gaussian Mixture Model (GMM) methods. A new unsupervised multi-person re-identification algorithm has been developed, which dynamically assigns labels to persons for recognition. Detection of persons that have ever been in motion but become stationary for a long time is a challenge in conventional motion-based foreground detection methods. The proposed algorithm overcomes this challenge using information from the bounding boxes, targeting persons from precedent frame where they last moved. Chan-Vese active contours method is used to get proper shape of persons and corresponding bounding boxes from the foreground extracted by using the traditional GMM method. The proper shape obtained from active contours method is used to minimize the area of background in the tracked bounding box, which increases the accuracy of person re identification. Parallel fusion of color moments and structure tensors are proposed to solve the problem of person re-identification. For re-identification, distinctive color features of the persons are extracted and stored on their first appearance in the field of view. In the subsequent appearances, their corresponding features are compared with the stored features using sum of absolute difference and are properly labelled based on the similarity measure. Experiments were conducted on IIT Patna database and our experimental study shows robustness in multi person tracking and re-identification. The result shows that this approach leads to improvement in multi-person tracking with 26.79% increase in accuracy compared to GMM and 85.33% correct re-identification of person.