| Abstract [eng] |
Lawn mowing robots use a boundary wire to define the working area of the lawn mowing robot. This wire is placed around the lawn area and acts as a measuring and positioning tool to let the robot know where to cut the grass. The robot follows the path of the wire and adjusts its direction accordingly. Although this method is effective in most cases, it has some drawbacks and limitations. First of all, the inflexibility due to the predefined wire path limits the robot's ability to adapt to changes in the layout of the lawn, for example by adding or removing flower beds, outdoor structures, or rearranging outdoor furniture. Another very relevant disadvantage is related to damage to objects on the lawn and wildlife safety issues. Some newer models of lawn mowing robots incorporate the use of intelligent sensors and technologies (e.g. GPS, LiDAR or image processing smart technologies) to avoid reliance on boundary markers. The aim of this work is to experimentally investigate and evaluate different deep learning architectures for environmental and obstacle recognition tasks in a field of robot mowers using the ADE20K dataset, and to develop a decision making module. The semantic segmentation models selected for the evaluation of the developed solution are FCN, LEDNet, ICNet, ContextNet, Deeplabv3+, FastSCNN and UNet. The experiments showed that the accuracy of the models increased as the number of classes decreased, and the best models - FCN and Deeplabv3+ - achieved as high as 92% segmentation accuracy. Due to the lack of examples for certain classes, a new dataset was constructed, upon which the trained models provided 98% accuracy in identifying the "cut surface", 85% accuracy in identifying the "driven surface" and 99% accuracy in identifying the "obstacle" class. Using these results, an algorithm was developed that assesses the situation in real time and performs decision-making, with the additional aim of reducing the robot's energy consumption. |
