SPIKING CONVOLUTIONAL RECURRENT NEURAL NETWORKS FOR SITUATIONAL AWARENESS (SCRNN) CLASSIFICATION, SEMATIC SEGMENTATION, AND TRACKING USING UAV EVENT
Abstract
Image recognition has become increasingly important in many industrial domains such as security systems, robotics, medical devices. Many developed algorithms have outperformed human performance in different image recognition tasks for example the success of Convolutional Neural Networks(CNN) in the 2012 ILSVRC image classification challenge. However, it remains a difficult task to extend the achievements in static image recognition to dynamic scene recognition such as action or moving objects. To remedy this situation, spike-domain sensing and processing is rapidly emerging as a new paradigm in learnable data-driven system design. The benefit of event-based brain like (Spiking or Neuromorphic) sensor on such motion recognition tasks is that it offers very high temporal resolution when a large fraction of scene changes, which can only be matched by a highspeed digital camera with the requirement of high power and significant resources. Airborne defense applications that involve autonomous intelligent systems and pervasive & ubiquitous computing, are often limited by the latency, power & computational costs associated with sensing and processing an ever increasing large volume of data. Spiking systems have the potential to combine lower latency with lower power through sparse encoding inspired by biological systems while achieving superhuman abilities in recognition and reaction. A novel Spiking Convolutional Recurrent Neural Network (SCRNN) model will be designed, implemented and tested during this research program. The SCRNN model will used for processing event data streams from a multiple UAVs for the problem of situational awareness, object classification, semantic segmentation and tracking. The relative SWaP, low latency and performance characteristics of the SCRNN model will be demonstrated on lab based and real datasets.
Document Details
- Document Type
- DoD Grant Award
- Publication Date
- Aug 11, 2021
- Source ID
- FA86552017037
Entities
People
- John Soraghan
Organizations
- Air Force Office of Scientific Research
- United States Air Force
- University of Strathclyde