Spatial segment-aware clustering based dynamic reliability threshold determination (SSC-DRTD) for unsupervised person re-identification

Person Re-Identification (re-ID) in a crowded multi-camera surveillance environment is a highly challenging task. The traditional benchmark datasets contain less number of occluded images due to the pre-planned setup and limited duration of the videos recorded. Unlike the traditional benchmark person re-ID datasets, real-world surveillance environment possess high static and dynamic occlusions. The analysis of different image segments captured in diverse environments by using a static reliability threshold leads to a poor matching accuracy. To resolve this issue of poor reliability threshold determination and to handle the occluded person re-ID images efficiently, we propose an unsupervised spatial segmented clustering model (SSC-DRTD) which determines a dynamic segment-wise reliability threshold. The unlabeled person re-ID images are segmented into k-parts to determine the segment-wise reliability threshold and the optimal number of segments for a given dataset. A cluster refinement strategy is proposed by incorporating the determined dynamic reliability threshold values to match the occluded noisy images with its appropriate ground truth identities for robust cluster formation. An improved rank evaluation has been performed on the benchmark person re-ID datasets such as DukeMTMC re-ID, Market1501, CUHK03, and MSMT17. The experimental results show the improved performance of our proposed SSC-DRTD model in handling occluded person re-ID images over the state-of-the-art unsupervised person re-ID methods. To further prove the efficiency of our proposed model, an exploratory analysis is performed by increasing the number of occluded query images to simulate the real-world surveillance scenario.

Automated summary

An unsupervised spatial segmented clustering model (SSC-DRTD) is proposed to handle occluded person re-identification images, showing improved performance over state-of-the-art methods on benchmark datasets.