Seeing triple
Compound eyes like those of insects may help drones find their way around
Jul 25th 2015
The eyes have it
GIVING sight to robots is an important goal, but a tricky one. Most attempts use cameras that produce the sort of image a human being is used to, and then apply computing power to simplify it (for example, by searching for the edges of objects) in ways that tell a robot what it needs to know (ie, do not blunder into that edge).
Dario Floreano of the Swiss Federal Institute of Technology, in Lausanne, has, however, taken a different approach. If simplicity is what is required, then simplicity is what needs to be supplied in the first place. Dr Floreano went to the natural world to look at a group of animals—the insects—that have often inspired robotmakers, with a view to copying the way that they organise vision. As he and his colleagues report in the Journal of the Royal Society Interface, the result is an artificial version of an insect’s compound eye.
Insect eyes are made of thousands of hexagonal columns called ommatidia, each of which focuses light through a lens down a transparent tube called a rhabdom to a set of photosensitive cells at the bottom. Such eyes do not have the resolving power of single-lens vertebrate eyes, but they are far better at detecting motion. This is a particularly valuable skill when much of the world views you as lunch, and any moving object may thus be a threat. But it is also true, from the point of view of a moving animal, that the environment itself appears to be moving. The motion-detecting skills of ommatidia can be used to analyse this apparent movement.
Each of Dr Floreano’s artificial ommatidia weighs a mere two milligrams. Each also has a tiny polymer lens that focuses light through a transparent glass stack (which stands in for the rhabdom), onto an array of three photodetectors arranged as an equilateral triangle. The stack has opaque walls, to stop light leaking into neighbouring stacks, so Dr Floreano’s artificial ommatidia can, like natural ones, be bundled together to form compound eyes.
Natural ommatidia detect movement through a phenomenon called “optic flow”. This is the pattern of apparent motion of objects in the visual field they are looking at, caused by the actual movement of the insect they are attached to. Objects close by, for example, will appear to move faster than more distant ones (think of the view from the window of a moving train). A predator will move at an angle to most other things in the field, so is easy to notice.
The idea of using optic flow for drone vision is not new, but Dr Floreano’s approach certainly is. Previous attempts have employed either standard (albeit miniature) cameras, which suffer from the same initial-complexity problem as those that work by looking for edges, or unsophisticated ommatidia that are able to measure optic flow in only one direction (left-right, up-down, near-far) at a time.
One of these old-fashioned ommatidia works by pairing up with a neighbour to analyse movement along the axis between the pair. Linking the outputs of several nearby pairs, aligned in different directions, permits the optic flow of one part of the visual field to be analysed completely. In Dr Floreano’s arrangement, by contrast, each of the vertices of the triangle of detectors inside a single ommatidium can pair with each of the others, to form three pairs in all, aligned at 120° to each other. A lone ommatidium is thus able to keep track (via some nifty computing) of everything going on, optic-flow-wise, in the part of the visual field it is pointing towards.
The team tested their new ommatidia by rotating them individually inside a room with patterned walls, and also by attaching an array of them to a wheeled platform that was sent travelling down a similarly patterned corridor. They measured the artificial eyes’ outputs, and compared them with a calculation of what they thought these should have been, given the patterns on the walls and the speed at which the ommatidia were moving. The two matched well: the ommatidia were behaving as they should. That suggests the signals they were producing could be relied on to steer at least an earthbound robot, once the algorithms needed to interpret these signals have been perfected.
Tests on drones will follow. These will be constrained by the weight of the computer needed for the processing. But if that can be miniaturised, the researchers will have taken a useful step towards robot aircraft that can see where they are going.
http://www.economist.com/news/scien...-may-help-drones-find-their-way-around-seeing
Compound eyes like those of insects may help drones find their way around
Jul 25th 2015
The eyes have it
GIVING sight to robots is an important goal, but a tricky one. Most attempts use cameras that produce the sort of image a human being is used to, and then apply computing power to simplify it (for example, by searching for the edges of objects) in ways that tell a robot what it needs to know (ie, do not blunder into that edge).
Dario Floreano of the Swiss Federal Institute of Technology, in Lausanne, has, however, taken a different approach. If simplicity is what is required, then simplicity is what needs to be supplied in the first place. Dr Floreano went to the natural world to look at a group of animals—the insects—that have often inspired robotmakers, with a view to copying the way that they organise vision. As he and his colleagues report in the Journal of the Royal Society Interface, the result is an artificial version of an insect’s compound eye.
Insect eyes are made of thousands of hexagonal columns called ommatidia, each of which focuses light through a lens down a transparent tube called a rhabdom to a set of photosensitive cells at the bottom. Such eyes do not have the resolving power of single-lens vertebrate eyes, but they are far better at detecting motion. This is a particularly valuable skill when much of the world views you as lunch, and any moving object may thus be a threat. But it is also true, from the point of view of a moving animal, that the environment itself appears to be moving. The motion-detecting skills of ommatidia can be used to analyse this apparent movement.
Each of Dr Floreano’s artificial ommatidia weighs a mere two milligrams. Each also has a tiny polymer lens that focuses light through a transparent glass stack (which stands in for the rhabdom), onto an array of three photodetectors arranged as an equilateral triangle. The stack has opaque walls, to stop light leaking into neighbouring stacks, so Dr Floreano’s artificial ommatidia can, like natural ones, be bundled together to form compound eyes.
Natural ommatidia detect movement through a phenomenon called “optic flow”. This is the pattern of apparent motion of objects in the visual field they are looking at, caused by the actual movement of the insect they are attached to. Objects close by, for example, will appear to move faster than more distant ones (think of the view from the window of a moving train). A predator will move at an angle to most other things in the field, so is easy to notice.
The idea of using optic flow for drone vision is not new, but Dr Floreano’s approach certainly is. Previous attempts have employed either standard (albeit miniature) cameras, which suffer from the same initial-complexity problem as those that work by looking for edges, or unsophisticated ommatidia that are able to measure optic flow in only one direction (left-right, up-down, near-far) at a time.
One of these old-fashioned ommatidia works by pairing up with a neighbour to analyse movement along the axis between the pair. Linking the outputs of several nearby pairs, aligned in different directions, permits the optic flow of one part of the visual field to be analysed completely. In Dr Floreano’s arrangement, by contrast, each of the vertices of the triangle of detectors inside a single ommatidium can pair with each of the others, to form three pairs in all, aligned at 120° to each other. A lone ommatidium is thus able to keep track (via some nifty computing) of everything going on, optic-flow-wise, in the part of the visual field it is pointing towards.
The team tested their new ommatidia by rotating them individually inside a room with patterned walls, and also by attaching an array of them to a wheeled platform that was sent travelling down a similarly patterned corridor. They measured the artificial eyes’ outputs, and compared them with a calculation of what they thought these should have been, given the patterns on the walls and the speed at which the ommatidia were moving. The two matched well: the ommatidia were behaving as they should. That suggests the signals they were producing could be relied on to steer at least an earthbound robot, once the algorithms needed to interpret these signals have been perfected.
Tests on drones will follow. These will be constrained by the weight of the computer needed for the processing. But if that can be miniaturised, the researchers will have taken a useful step towards robot aircraft that can see where they are going.
http://www.economist.com/news/scien...-may-help-drones-find-their-way-around-seeing
