Spiking Neural Networks for Advance Signal Processing Applications
ISL is pursuing research into alternative massively parallel architectures using coupled nonlinear circuits to bypass fundamental processing speed and power limitations inherent with conventional digital circuitry (e.g. parasitic capacitance and transistor switching rates). Multi-layer networks can be designed to recognize spatial and temporal correlations in analog signals and thereby perform high-speed pattern recognition. The archetype nonlinear circuit is the artificial spiking neuron (or integrate and fire circuit). Multi-layer parallel networks of spiking neurons with inhibitory and excitatory connections are ideal for performing high-speed low power computation. Processing occurs in a hybrid digital/analog environment with the amplitude of spikes represented digitally and the timing of spikes indicating temporal correlations. In this environment, practical spiking neural networks exhibit 3 to 4 orders of magnitude less power consumption than their digital counterparts.
Concepts include spiking neural networks implemented to perform analog to digital conversion (ADC) at high-speed (multiple GHz) with large dynamic ranges (and spurious free dynamic ranges). These advances in ADC technology will enable corresponding performance improvements in modern radar systems including direct sampling of RF signals (bandpass architectures) and detection of reduced signature targets in heavy clutter and electronic counter measure environments.
System Concept Development and Analysis
System concept development and analysis at ISL is based on a strong physics-based domain expertise in propagation, electromagnetic analysis, and other relevant phenomenology. This enables solutions that account for the real-world environment in which the system will operate as well as realistic hardware limitations. All aspects of ISL development (including algorithm development and implementation and modeling and simulation) are accomplished within this systems framework resulting in a more robust solution set for our customers.
Advanced Computing Platforms
ISL has software expertise and tools for exploiting emerging processing technologies such as graphical processing units (GPU) to meet the challenging computing requirements of modern sensor systems and simulations. ISL has demonstrated the use GPUs and other multi-core processors to speed up radar processing algorithms including space-time adaptive processing (STAP). ISL has established test bed hardware and software required for rapid prototyping of computational adjunct concepts.
Persistent Surveillance Exploitation
Detection, tracking, and identification of targets in a complex environment with a high density of background traffic is very difficult. The number of detections and tracks generated by the background traffic can easily overwhelm an operator. Persistent surveillance sensors that can cover a wide area over a period of time can enable a unique capability to separate unusual or suspicious activity from the normal background activity. ISL has developed innovative processing techniques to quantify the normal activity in a given area and exploit this knowledge to identify suspicious activity, cue an operator with a high confidence detection of targets of interest, and to efficiently manage sensor resources.
Precision Geolocation & Georegistration
A wide variety of small, low cost, unmanned air vehicles (UAVs) carrying video sensors are increasingly deployed on today’s battlefield for intelligence, surveillance and reconnaissance (ISR) and combat missions. These video sensors provide a considerable amount of data for exploitation, suitable for various types of assisted and autonomous processing, including target recognition, tracking, contextual scene analysis, aimpoint generation as well as forensic analysis, behavioral characteristics and inference. In general, sensor model-based methods provide the most accurate approach to georegistration. However, smaller UAVs may not have well calibrated sensors or their existing sensor models and sensor pointing directions may not be available. These limitations make associating accurate positioning information with the video frames problematic using sensor model-based techniques.
ISL is developing a capability to automatically produce accurate geopositioning information from UAV collected video in real time using advanced image registration techniques. Since this approach does not require a sensor model, it is particularly well suited for applications in which UAV video data has poor or unknown sensor metadata. A high-accuracy location estimate for each pixel in the image can be generated quickly by registration with a geo-spatial reference database. ISL has developed the modified log-polar transformation (LPT) registration technique to perform this registration with high-precision and modest computational cost. This advanced image registration technique can also be applied to detection of moving targets in video data and non-GPS navigation.
