Programmable Neuromorphic Microchip for Accelerating Dismount Identification in WFOV Video

Period of Performance: 08/27/2014 - 11/29/2016

$749K

Phase 2 SBIR

Recipient Firm

Isocline Engineering
1301 Beal ave
Ann Arbor, MI 48109
Principal Investigator

Abstract

ABSTRACT: Wide field of view (WFOV) video capability on unmanned aerial systems (UAS) has outpaced data transmission capability. Identifying regions of interest (often performed through object recognition) enables selective video compression and maximizes the useful content in data transmission. Neuromorphic classifiers currently hold the records for greatest accuracy of object recognition and speech recognition, but are too computationally intensive for deployment in autonomous systems. This proposal describes a programmable microchip for energy-efficient acceleration of neuromorphic classifiers. Neuromorphic classifiers interpret sensor data much like the human brain does, allowing them to efficiently recognize dismounts, equipment, buildings, terrains, and other interesting features in video. The proposed platform will increase the capabilities of UAS while greatly reducing the form factor and power requirements of the control systems. In our Phase I effort we developed a mixed-signal circuitry that showed a 46-360x improvement in power efficiency, and a 700-9200x improvement in performance per volume, in comparison to a digital ASIC implementation or FPGA implementation, respectively. In this Phase II effort we will fabricate a prototype system that meets the target requirements of the solicitation. BENEFIT: The proposed system can function as a general purpose data classifier with little modification. Many commercial electronics are looking for new ways of interacting with the outside world that can benefit through this technology. In each of these applications the proposed system will be able to perform the task with much greater accuracy and performance with a lower power budget than other solutions. Military applications include object recognition for autonomous systems, navigation for autonomous systems through optical flow, and voice and gesture recognition for human interfaces. Potential consumer applications include medical devices that include data classifiers for detecting heart attacks or seizures, speech and object recognition for handheld devices like cellphones, and human interfaces for automobiles.