Robot Global Relocalization Based on Multi-sensor Data Fusion

Abstract

In order to solve the problems of low localizable accuracy, slipping and “kidnapping”, which may lead to get “ LOST” in robot localization, a multi-sensor data fusion localization method that integrates the wheel encoder, the Inertial Measurement Unit (IMU), the optical flow s ensor a nd t he laser is proposed. Firstly, the Federal Kalman Filter (FKF) is used to fuse multi-sensor data information to obtain more accurate localization; Secondly, when the robot loses its global localization, the optical flow s ensor, t he r angefinder an d IM U ar e fu sed to obtain the localization as a priori for the prediction step in the resampling of Adaptive MonteCarlo Localization (AMCL). Finally, the Conditional Variational Autoencoder (CVAE) is used for training to further optimize the priori localization. The algorithm without correct initial values is converted into an algorithm with fuzzy localization. Experimental results based on real scenarios showed that the multi-sensor data fusion not only helped to obtain more accurate and stable location, but also significantly reduced the “LOST” problem in case of the robot being kidnapped compared to the pre-optimized AMCL algorithm.