STATE ESTIMATION IN MULTI-SENSOR FUSION NAVIGATION: EQUIVALENCE ANALYSIS ON FILTERING AND OPTIMIZATION

Abstract. Integration of information from multi-sensor can provide navigation systems with reliable estimates of their own states, which overcomes shortage of standalone sensors. State estimation approaches, which can be categorized into filtering-based and optimization-based methods, provide the means to fuse information from various sensors with different principles to estimate the system's position, orientation, and other navigation parameters accurately. Recent researches have shown that optimization-based frameworks outperform filtering-based ones in terms of accuracy. However, both methods are based on maximum likelihood estimation (MLE) and, assuming Gaussian noise, should be theoretically equivalent. In this paper, we comprehensively and theoretically analyse the differences between the two methods, including algorithms and strategies. Our simulated experiments based on visual odometry (VO) indicate that filtering-based approaches are equal to optimization-based ones in accuracy when employing the same strategies and under the premise of the same measurements and observation model. Therefore, future research on sensor-fusion navigation problems should concentrate on strategies rather than state estimation methods.