Reliable and accurate estimation of traffic states play an important role in traffic management and traffic theory development, which therefore has significant social and scientific relevance. Data for traffic estimation normally from different sources and have different types, characteristics, etc, so data fusion techniques are used. Traffic state estimation involves fusion of data from different sources. Therefore, data fusion techniques are used. The subject of this thesis is about traffic data fusion and the main objective is to propose more efficient approach and algorithms to accomplish traffic data fusion. Different types of data have different characteristics and accuracy, which causes challenges. Let us take data from loop detectors as an example. Loop detectors can provide the speed measures only at certain points on a road (local speeds). The speed measures from loop detectors have structural deviation from the ground-truth speeds based on Edie’s definition. The deviation is relatively bigger when the speed is lower. This deviation can lead to the 100% error in density estimation. Similary, some other types of data, e.g. floating car data, camera data, etc, have their particular characteristics. It is a challenge to fuse all kinds of data with different characteristics, semantics, resolution, accuracy and reliability. Although the previous methods have already solved quite a few traffic data fusion issues, yet there are quite a few challenges left. For example, due to spatio-temporal alignment problem, Kalman filter, the most commonly used assimilation techniques for data fusion can not be well used to fuse travel times and local data. Majority of these methods need model calibration that is made through biased data e.g. biased loop speeds, so these methods are not effective in removing the structural bias in data. In sum, previous data fusion techniques normally involves quite a few assumptions, but they may not fuse many types of data or give reliable results. In order to fuse more types of data and give more accurate results, we propose a new approach. This approach is called ‘Data-Data Consistency’ Approach. It still needs traffic models, but these models are simply based on some basic physical laws and very few assumptions. Based on this data-data consistency paradigm, we develop four methods for traffic data fusion. The first proposed algorithm is called PISCIT which is able to fuse traffic speeds from local detectors such as inductive loops with individual travel times measured by AVI systems. The second is called TravRes which is able to accurately reconstruct high resolution time-space speeds from floating car data (FCD). It achieves this by reconstructing the (unobserved) probe vehicle trajectories between polling time instants, until the resulting time-space speed map is consistent (enough) with all probe vehicle reports. The above-mentioned two algorithms are concerned with the low-resolution travel time data (low polling rates). The third is called FlowRes. It deals with another type of data, data which may not only have low time-resolution but also have quite low position resolution. Such data cannot pinpoint the accurate positions of vehicles but can only give some location-specific information when and where the vehicles are located at the segment or cell level.This algorithm corrects strongly biased prior speed measurements and reduces the impact of random errors. it can be easily extended to fit in network wide traffic speed estimation. The fourth algorithm is called ITSF, which is able to fuse traffic flow, local speeds and travel times all together. It uses extra data source: traffic flow. As a result, more accurate and reliable estimation is achieved compared to the first two algorithms.