tailieunhanh - Machine Learning and Robot Perception - Bruno Apolloni et al (Eds) Part 4

Tham khảo tài liệu 'machine learning and robot perception - bruno apolloni et al (eds) part 4', kỹ thuật - công nghệ, cơ khí - chế tạo máy phục vụ nhu cầu học tập, nghiên cứu và làm việc hiệu quả | 68 M. Yeasin and R. Sharma Space-variant Vision Sensor With the introduction of biological concepts the space-variant vision architecture and issues related to this is gaining momentum especially in the field of robotic vision and image communication. In designing a visual sensor for an autonomous robot or any other visual system in which the constraints of performance size weight data reduction and cost must be jointly optimized five main requirements are imposed 1 a high resolution fovea for obtaining details in the region of interest 2 a wide field of view useful for many tasks for example interest point determination 3 fast response time5 4 smooth variation of resolution across the visual work space and finally 5 the cost size and performance6 of the sensor. To develop a space-variant vision sensor several attempts to combine peripheral and foveal vision have already been made in the past decades. For example space-variant image sampling 55 and combination of wide and tele cameras 56 57 but such methods are not discussed in this chapter as they do not fit in to the context of our discussion. Studies reveal that there are mainly two types of space-variant sensors available and the clarification of this issue will go a long way towards clarifying several basic issues. First one could work in cortical plane which has fundamentally different geometry than the retina but in which the size of the pixels increases towards the periphery. Second one in which the image geometry is still Cartesian but retains the same space-variance in the pixel structure. Successful efforts in developing space-variant sensors are summarized in subsequent subsections. Specially Designed Lens This approach combines a wide field of view and a high resolution fovea by specially designed lens. The purely optical properties of this method avoid most of the problems involved in space-varying sensor design and implementation . co-axial parallelism continuity hardware redundancy and