This paper explains the increase in localization system accuracy of the adaptive Monte Carlo localization (AMCL) in robots utilizing a convolutional neural network (CNN). The localization system in robots is defined as the position recognition process of robots within their working environment. This system is essential as it allows robots to navigate and map efficiently and accurately. Without appropriate localization, robots cannot operate effectively and can encounter troubles such as losing direction or bumping into objects. AMCL is a popular localization system and is widely applied in robots. This method utilizes the changes in the robots’ position and light detection and ranging (LiDAR) sensor reading as input. Reading of robot position changes is susceptible to error due to slips or wheel deformations. The inaccuracy of reading the robots’ position change results in the inaccuracy of the robots’ position prediction by AMCL, so improvements are required. Novelty in this paper includes providing compensation values from AMCL results for the error to be small. These compensation values were obtained from the CNN training results; hence, the proposed method was dubbed AMCL+CNN. Inputs given to the CNN were the changes in wheel odometry values and distance reading by the LiDAR sensor. CNN outputs were compared to the target data in the form of the robots’ actual position from observation results. Network training was conducted for as many as 200 epochs to achieve the lowest validation loss. Testing was done on a robot installed with a robot operating system (ROS). Training and testing datasets were obtained from rosbag data when the robot traversed the testing area. In straight and turn scenarios, obtained AMCL+CNN algorithms had fewer errors than the regular AMCL and Monte Carlo localization (MCL). Results obtained are also superior in terms of positional error metrics when compared to several other comparison methods.
Copyrights © 2023