A large collections of prior techniques proposed in WIFI indoor localization using received signal strength fingerprint. However, the practical WIFI localization system has not been used in large scale environment. Little prior research and industry systems work the large scale implementation due to lack of efficient way to collect and construct fingerprint database. The accuracy of localization is also a challenge problem which limits the large scale usage of indoor localization system for the small scale of indoor space needs more accurate than outdoors, and the variance of WIFI signals change heavily causes incorrect location estimation. Therefore, a robust indoor localization method that both consider the practical WIFI data collection reduction and to improve the positionin...[ Read more ]

A large collections of prior techniques proposed in WIFI indoor localization using received signal strength fingerprint. However, the practical WIFI localization system has not been used in large scale environment. Little prior research and industry systems work the large scale implementation due to lack of efficient way to collect and construct fingerprint database. The accuracy of localization is also a challenge problem which limits the large scale usage of indoor localization system for the small scale of indoor space needs more accurate than outdoors, and the variance of WIFI signals change heavily causes incorrect location estimation. Therefore, a robust indoor localization method that both consider the practical WIFI data collection reduction and to improve the positioning accuracy by reducing data noise is needed. The goal of this research work is to design a crowdsourcing mobile application, and to collect WIFI data with no labor consuming, and proposed a better algorithm in order to deal with the noisy crowdsourcing data.

In this thesis, we present our multi-task learning based deep Gaussian process model to address the challenging issues in nowadays indoor localization problem. We introduced a novel indoor localization method without any data collection labor, which is potentially suitable for large scale implementation. We first proposed a framework of multi-task learning in deep Gaussian process. Deep Gaussian process is utilized in order to deal with the big and noise label issues, and the multi-task is capitalized to deal with differences across devices. In the multi-task learning, we designed two different parameter sharing methods in order to transfer knowledge between tasks. Our experimental evaluation shows that our algorithm outperforms any other state-of-art methods.