Abstract: Accurate, continuous and reliable location information is the basic condition for in-vehicle navigation applications. Under the premise of not adding other sensors, integrating GNSS, MEMS, on-board CAN sensors and vehicle motion constraint information is the most practical and low-cost vehicle multi-fusion navigation scheme. In the vehicle motion constraints, reasonable configuration of the relevant parameters is the key to making the constraints work fully. Thus, focusing on the vehicle non-integrity constraints, this article uses multiple regression and deep learning methods to build a new vehicle motion constraint model with parameter autonomous learning. Moreover, a new idea of directly learning lateral/vertical velocity parameters in the observation domain is proposed, which has better constraint effect than the old variance domain parameter adjustment method. The experiments show that compared with the traditional method of adjusting parameters in the variance domain, the new model with parameter autonomous learning in the observation domain has a significant improvement in accuracy. The inertial estimation error using the multivariate regression models is reduced by 69.6%~81.2% in the horizontal position, while the use of deep learning is reduced by 60.0%~77.3%. At the same time, the horizontal relative positioning accuracy is improved by 75.2% and 65.0% respectively, the new model can effectively improve the maintenance ability of vehicle positioning accuracy when GNSS failure.

Key words: vehicle navigation, multi-source fusion, parameter autonomous learning, vehicle motion constraint, non-holonomic constraint