A Reliable Position Estimation Methodology Based on Multi-Source Information for Intelligent Vehicles in Unknown Environment.

Accurate position is important for intelligent vehicles especially in unknown environments without prior information. Light Detection and Ranging (LiDAR) has shown great potential in the application of intelligent vehicles, especially when satellites fail. However, the following should be considered to further improve its positioning performance. Firstly, the LiDAR positioning accuracy is influenced by the initial iteration value. Secondly, the cumulative error of LiDAR positioning cannot be ignored due to its incremental positioning method. To circumvent aforementioned problems, a reliable position estimation methodology based on multi-source information is proposed. Initially, a deep neural network based on multi view features multi-head self-attention fusion is designed to estimate the relative heading value to offer initial value for LiDAR positioning. Then, the cooperative positioning multilayer perceptron (CPMLP) is proposed to obtain reliable cooperative position observation just by one road side unit to suppress the cumulative error especially in the environment without prior information. Lastly, the probabilistic graphical model is used to fuse multi-source information. Extensive dataset and real experiments show the effectiveness of our methodology.