Abstract:
For video analytics and surveillance applications, person re-identification across multiple camera views remains an open problem. The challenge of being able to determine that two images of people are the same person based solely on their appearance can be difficult, even for human observers. Many recent re-identification methods use deep learning with supervised learning to discriminate between the identity classes. However, the requisite training data is generally not available in real-world scenarios. In this paper, we compare a number of fast classification methods for the purposes of re-identification, taking extracted and pre-processed feature vectors and classifying them into identity classes, focusing on one-shot and unsupervised learning algorithms. We present two novel one-shot learning methods, including Sequential K-means, a computationally efficient algorithm with competitive accuracy. We demonstrate this on an indoor person tracking dataset, and discuss parameter tuning in order to further improve the accuracy of the algorithm.