Identity retention of multiple objects under extreme occlusion scenarios using feature descriptors

Abstract

Identity assignment and retention needs multiple object detection and tracking. It plays a vital role in behavior analysis and gait recognition. The objective of Multiple Object Tracking (MOT) is to detect, track and retain identities from an image sequence. An occlusion is a major resistance in identity retention. It is a challenging task to handle occlusion while tracking varying number of person in the complex scene using a monocular camera. In MOT, occlusion remains a challenging task in real world applications. This paper uses Gaussian Mixture Model (GMM) and Hungarian Assignment (HA) for person detection and tracking. We propose an identity retention algorithm using Rotation Scale and Translation (RST) invariant feature descriptors. In addition, a segmentation based optimum demerge handling algorithm is proposed to retain proper identities under occlusion. The proposed approach is evaluated on a standard surveillance dataset sequences and it achieves 97 % object detection accuracy and 85% tracking accuracy for PETS-S2.L1 sequence and 69.7% accuracy as well as 72.3% precision for Town Centre Sequence.