Vision and Machine Learning in Autonomous Lawn Mowers

The progress in the field of Computer Vision, especially thanks to the advent of neural networks and ever more performing graphics cards, have reached previously unimaginable levels. Because of this, the industry is integrating these new techniques with already existing technologies, ranging from more diverse fields, such as medicine, video surveillance, quality control and so on. The objective of this thesis is to analyze which are the vision techniques that are best suited to a specific field, the autonomous lawnmower robotics. The obstacle avoidance process of this category of robots can be facilitated through various methods, but during this study the calculation of depth maps from a stereo camera and grass non-grass segmentation from images were taken into consideration, with a focus on obtaining real-time performances. For this reason, the text is divided into two parts: the first part aims to be a comparison between some convolutional neural networks suitable for semantic segmentation and verify how reliable these are for the purpose. There are investigated also some common patterns in CNN's architectures to make them efficient as possible. The second part aims to analyze the possibilities of obtaining how far the obstacles are from the robot using classical algorithms for calculating disparity maps. The differences in terms of quality and time for computation compared to new techniques based on deep learning are also analyzed. For both parties, this study reveals that both approaches, with some precautions, can represent a valid aid for the future development of autonomous lawnmowers. Classical methods and methods based on machine learning both have advantages and disadvantages, so making them coexist could be a solution that comes close to the best one. On this starting point it is possible to make these robots increasingly autonomous and reliable. Investigating new methods and taking inspiration from what has been done, trying to generalize, then making them as robust as possible to all possible scenarios is the main challenge that will determine the applicability in the industry.