Abstract
Accurate localization is a critical requirement for the successful deployment of mobile robots in indoor environments, particularly those characterized by symmetrical structures and features. Symmetrical indoor environments pose unique localization challenges due to the presence of repeated patterns and structures that can confound traditional localization methods. In such environments, accurately estimating the robot's pose relative to a global reference frame becomes increasingly challenging, leading to potential errors and inefficiencies in robot navigation and task execution. This paper presents an improved deep learning-based Monte Carlo localization (DMCL) framework for global localization of a mobile robot in symmetrical indoor environment using only 2D lidar. We first, converted 2D laser data to single channel 2D projected image and an occupancy grid. This 2D projected image is used to train the neural network to regress the 3DOF of robot. Finally, we integrated this trained neural network which estimate the robot pose in environment with MCL in the weight updating stage. The performance of both Monte Carlo Localization (MCL) and in DMCL methods in symmetrical indoor environment is investigated through extensive simulation studies. Verifying the effectiveness of the proposed method our network is able to obtain position accuracy of 0.15m and scene classification accuracy to 99% in simulation.