One of my research focus is the design of high speed bio-inspired vision sensors. Conventional frame-based vision systems are limited in speed. Changes cannot be detected when one pixel is being scanned. They are also limited in resolution. If the number of pixels of the pixel array is increased, the time to scan all the pixels and generate one frame will be higher.
With the bio-inspired vision systems, we leave the notion of frame. Information is transmitted in a continuos way (only if there is a relevant change to transmit) and pixels do not need to be scanned. This made them inherently faster. We can also increase arbitrarily the number of pixels without reducing the pixel speed. The only limitation is the speed of arbitration system and the capacity of the systems that have to process the information generated by the pixel array.
We designed and created a vision sensor capable of detecting movement. It had an event-based output. The motion detection is one of the most important tasks performed by the human retina. It is specially important to obtain information about our environment. I was in charge of all the experimental setup of this project collaborating with my former advisors Bernabe Linares and Teresa Serrano. This project was done at the Seville Microelectronics Institute (IMSE).
In this project, we improved some features of the previous Dynamic Vision Sensor (DVS) designed by Lichtsteiner et al. To do this, we proposed a new topology for the photoreceptor to increase the pixel bandwidth. We also increased the pixel gain to reduce the latency time response. We did this simply adding two small amplifiers at the output of the receptor. In the previous the design, the gain was equal to the ratio between the capacitors. For this reason, the pixel size was higher. We developed a pixel with more speed and reduced size just using amplifiers. The only drawback of this approach was an increment of the power consumption.
.
The system was capable of detecting very fast moving objects. To generate points rotating with a controlled speed, we used one analog oscilloscope working on XY mode. We can see above one snapshot of the input stimulus taken with an conventional camera (left side) and the spatio-temporal representation of the data generated by the DVS (right side). The rotating dot frequency was 500Hz (30,000 rpm). We represent the negative events (transitions from bright to dark) in red, and the positive events (transitions from dark to bright) in blue. The sensor can detect the sign of the temporal contrast. We were able to detect the movement of rotating points with speeds of 10KHz (600,000 rpm!!).
In the clip showed below, we reproduce the recorded data using the jAER interface. The stimulus was a rotating dot with a frequency of 5KHz (300,000 rpm). We start reproducing the recorded data at normal speed. After, we reproduce is slowly to see the trajectory of the dot.
In the video displayed below, you can see very fast moving cards. We recorded the events generated by our sensor. First we reproduce the recorded video at normal speed. Then we reproduced it slowly. We can identify the different cards. The period between two different cards was approximately 1ms.
Project Publications
Back
