Sensor design with embedded low power neural networks for edge computer vision

Recent advances in neural network design will be leveraged for this embedded-ML sensor project. Ultra-low power fast acting neural networks can be built using neurons which process information inherent in a signal consisting of a train of spikes over time. These networks can be trained to recognise the 'temporal fingerprint' of a particular subject or action occurring within the field of view of a sensor array.

Such an array could consist of passive light detectors or active detectors, such as SPADs, in tandem with a laser illumination source. The latter allows for distance or time of arrival as well as intensity information to be captured and processed by the network.

Such a low-power network could form a 'trigger' or 'wake-up' layer for a more accurate and power consuming network which performs detailed object detection/classification or carries out actions. This 'wake-up' layer would allow for the main network to be left dormant, only activating when the need arises and thus minimizing the overall system power.

Low voltage operation and power-saving techniques will be investigated. Methods of massively parallel connection between imager and network will be researched.