دانلود کتاب چشم انداز رباتیک – الگوریتم های بنیادی در متلب

Digital images are everywhere – most of us have collections of thousands of digital images captured on a variety of devices. Each image is rich in information about the scene it recorded – implicit in every pixel is information about the 3D structure of the scene, the materials which comprise it and the lighting conditions at the time. In our everyday life we use our sense of vision to help us perform a huge variety of tasks: recognizing people, reading text, manipulating objects and navigating. Some simple visual skills can now be performed by machines, for example, cameras can tell if we’re smiling, social media platforms and customs barriers can recognize individual faces; and we can search our online photo albums using text keywords. Robots need mastery of visual skills if they are to meet their potential as helpful and ubiquitous machines in our everyday lives. Robots are data-driven machines – they acquire data, process it and take action based on it. The data comes from many sensors including cameras. An image comprises millions of pixels and the amount of data that needs to be processed, in realtime, is massive. Progress robotic vision has been, and continues to be, driven by more effective ways to process that data. This has been achieved through new and more effi cient algorithms, as well as the dramatic increase in computational power that follows Moore’s law.⊳ When I started in robotics and vision in the mid 1980s, see Fig. 0.1, the IBM PC had been recently released – it had a 4.77 MHz 16-bit microprocessor and 16 kbytes (expandable to 256 k) of memory. Over the intervening 30+ years computing power has perhaps doubled 20 times which is an increase by a factor of over one million. In the last 5 years, graphical processing units, or GPUs, have come to the fore and powered a new wave of data-driven algorithms most notably deep learning⊳. Over the fairly recent history of robotic vision a very large body of algorithms has been developed to effi ciently solve large-scale problems in scene understanding⊳, motion estimation, and 3D reconstruction – a signifi cant, tangible, and collective achievement of the research community. However its sheer size and complexity presents a very real barrier to somebody new entering the fi eld. Given so many algorithms from which to choose, a real and important question is: What is the right algorithm for my particular problem? One strategy would be to try a few different algorithms and see which works best for the problem at hand, but this is not trivial and leads to the next question: How can I evaluate algorithm X on my own data without spending days coding and debugging it from the original research papers?