The neuromorphic stereo sensor is the realization of our effort in combining stereo vision methods with neuromorphic optical sensing. It consists of two 128 x 128 pixel sensor chips mounted on a board with a separation between the two chips of about 13 cm and a single processing unit.

Using the optical transient sensors helps to minimize the processing power needed, which brings the possibility to do distance calculation on-board in a single signal processing unit. Furthermore, there are enough free capabilities to run an application in the sensor system, too. As the chip detects moving objects only, no resources are spent to do unnecessary calculations on background objects. The neuromorphic stereo sensor system offers the following advantages:

• Real time embedded stereo processing with up to 100 pseudo frames per second. Simple applications can also be done on-board.
• Adding new information to the compact address-event format. The 2D sensor chip output becomes a continuous 3D output stream available for further data analyses respectively applications. The result can also be represented as a distance map showing moving objects.
• Advantage of all the stereo vision properties, i.e. insensitivity against shadows or difficult illumination conditions.
• Inherently detects moving objects which is equal to the interesting part in a scene for most applications. Background objects are eliminated.

An already realised example of an application is the counting of people. Our smart eye UCOS – Universal Counting Sensor has been engineered to do this.

The image below shows example data for persons crossing the sensor field of view (overhead) both on the two individual sensors (left) and the stereo reconstruction (right, colours from blue to red represent depth).

A particularly useful characteristic of the neuromorphic stereo sensor in this application is that it detects shapes only, but does not provide images in which persons could be identified. Thus, no privacy issues can occur.