Unmanned Systems Technology 021 | Robot Aviation FX450 l Imaging Sensors focus l UAVs Insight l Liquid-Piston X-Mini l Riptide l Eurosatory 2018 show report l Zipline l Electric Motors focus l ASTS show report

36 Focus | Imaging charge in certain conditions. This is an architectural change rather than just thickening the epi layer of the pixel to collect more photons as they pass through. Sensor makers say it is not possible to achieve 120 dB by thickening the epi, but there are additional steps in the pixel that help increase the dynamic range. The pixel size for most CMOS sensors for automotive applications these days is in the 3-4 micron range, but it can be as small as 1 micron to fit more pixels onto a chip to deliver a higher resolution image. Alongside the pixel size, the challenge is to minimise the current that builds up in the sensor even when there are no photons (called the dark current) and the electrical noise, as well as reducing the overall power consumption. The power per area remains constant regardless of the number of pixels or the resolution, but extra heat creates more dark current, so reducing the power consumption that leads to heating effects improves the performance. Reducing the power also reduces the need for thicker wires in the car, reducing the weight and cost of the cable harness. Volume manufacturing One of the largest sensor makers started out building them on 200 mm diameter CMOS silicon wafers at 110 and 90 nm geometries. It is now using a 65 nm process on a 300 mm wafer, which increases the number of sensors produced by more than 50%. Being able to do this is more complex than for standard CMOS chips, as wafers are stacked on top of each other to get the final sensor. One wafer has the pixels, while another has the analogue and digital circuitry for collecting the charge and providing the output. The pixel wafer is mounted upside down on the other wafer (metal layer to metal layer) and the pixel wafer is then ground down to a thickness of 4-6 microns. This allows light through the back of the wafer, and is called backside illumination (BI), and then RGB filters and micro-lenses can be added over each pixel to focus the light onto the active area. This type of BI structure is not possible with CCD sensors as they use a different pixel structure. OEMs have different ideas on how many sensors and what resolution are necessary in a driverless car, varying from 1 or 2 MP up to 8 or even 12 MP. The current ‘sweet spot’ is 7-8 MP to balance the image quality, cost and power consumption. Car makers are looking for redundancy in the image systems, so front-facing image units are using multiple high- resolution sensors and multiple processing in parallel. These are often tri-focal cameras with three sensors of different resolutions looking at different fields of view (FoV) at different distances. For example, the 7 MP sensor will be looking at 200 m with a narrow, 6 º FoV for long-range detection, an intermediate FoV for the mid-range, say 80 m, and a wide FoV to see things at the side of road within 10 m. That gives redundancy, so if one camera fails then the car can only see out to 80 m instead of 200 m, but that is good enough to keep going. This gives what is called ‘graceful degradation’. For the designs coming to market in 2021-22 for driver assistance and Level 3 autonomy, most car makers have standardised on 7 MP sensors with a wide field of view. 7 MP is seen as a good size for distance and the size of objects likely to be detected, such as pedestrians, cyclists and road signs. Higher resolutions than that would August/September 2018 | Unmanned Systems Technology The AR0430 is a backside illuminated 4 MP CMOS image sensor that can capture 120 fps and a depth map simultaneously from a single device (Courtesy of On Semiconductor) For the designs coming to market in 2021-22, most car makers have standardised on 7 MP sensors which have a wide field of view