Person re-identification with multiple similarity probabilities using deep metric learning for efficient smart security applications

Surveillance video analysis plays a vital role in the daily operations of smart cities, which increasingly relies on person re-identification technology to sustain smart security applications. However, research challenges of re-identification remain especially in terms of recognizing the different appearances of the same person in a harsh real-world environment: (1) the adaptability of the selected features to the dynamic environment cannot be guaranteed, and (2) existing methods rooted from metric learning aim to find a single metric function, and they lack the ability to measure the different appearances of the same person. To address these problems, this study proposes a multiple deep metric learning method empowered by the functionality of person similarity probability measurement. The proposed method exploits multiple stacked auto-encoder networks and classification networks to quantify pedestrians’ similarity relations. The stacked auto-encoder networks directly recognize persons from surveillance images at the pixel level. The classification networks are equipped with the Softmax regression models and produce multiple similarity probabilities to characterize different appearances belonging to the same person. An Adaboost-like model is designed to fuse the probabilities corresponding to multiple metrics, which ensures a high accuracy of recognition. Experimental results on two public datasets (VIPeR and CUHK-01) indicate that the proposed method outperforms existing algorithms by 2%–10% at rank 1. Based on the similarity probabilities learned by the proposed model, the algorithm for matching the person pair can achieve a time complexity as low as O(n), which can be deployed at a large scale on the distributed intelligent surveillance network, with each node maintaining limited computing capabilities.