High-Throughput Computer Vision/Etegent Technologies, LTD

The Etegent team proposes to design and demonstrate the feasibility of a hardware processor system capable of supporting computer vision (CV) object detection on tens of gigapixels per second in SWaP (Size, Weight and Power) limited environments. Specifically, this proposal will design and implement a modular and composable parallel computing software framework and prototype hardware system enabling the rapid transition of computer vision algorithms types from a development environment to a high performance, low-power, parallel computing platform. The team will research, augment and enhance existing inferencing tools for faster transition from development to deployment of more sophisticated, heterogeneous neural network topologies. As proof of feasibility, this effort will transform and implement several complete object detection neural networks using the developed tools and prototype hardware. Additionally, the team will characterize the implemented approach to measure and verify: 1) system throughput; 2) hardware power consumption; 3) implementation time required to transition networks to the system; and 4) performance of system operating with various sensor data formats. Results of this effort will demonstrate and deliver the software tools and applicable hardware design to achieve high throughput computer vision in SWaP limited environments.