Cross-Domain Adaptation Through Adversarial Teaching for Unmanned Ariel Vehicles in Smart Cities

Person re-identification has been an important issue of surveillance systems in smart cities. However, this requires huge datasets to supervise deep learning models for accurately identifying and tracking people in smart cities. In this paper, we propose adversarial learning in integration with a mutual mean teaching model known as adversarial teaching to take the issue of domain adaptation. Initially, they learn a source dataset in a supervised way. This model information is then transferred to another model through unsupervised learning and pseudo-labels are generated by generative adversarial learning. The multi-class invariance is handled by adversarial learning. The performance of the proposed model is evaluated on two different datasets source and target datasets. Multiple experiments on Duke and Market datasets indicate higher accuracy of F-1 score on both datasets compared to previous methods, demonstrating the superior performance of the proposed method.