NOVATEUR RESEARCH SOLUTIONS LLC — Department of Defense SBIR Phase I: AF161-138

Description

ABSTRACT: This SBIR Phase I project proposes a hierarchical visual processing system and attention-based mechanisms based on novel 3D convolutional neural networks to enable robust and efficient detection and classification of objects in 3D LIDAR imagery. Our approach models the ventral pathway of the visual cortex and addresses many limitations of the state-of-the-art solutions that perform exhaustive search for the targets of interest and rely on pre-designed features. Instead, the proposed framework uses bottom-up and top-down attentional mechanisms to improve the efficiency of the framework and enable incorporation of prior models and available context for efficient performance. In addition, instead of using hard-coded features, the proposed approach is capable of learning hierarchical invariant 3D feature representations that are object class independent from training datasets. The Phase I effort will include; development of biologically inspired 3D deep learning framework for 3D LIDAR data, fusion-of-attention mechanisms to identify salient regions, implementation of the software as open-source library, quantitative and qualitative evaluation of the proposed technologies, and demonstration of proof of concept using real-world data from multiple use-cases.; BENEFIT: Autonomous unmanned vehicles (UxVs) have proven to be critical assets for intelligence, surveillance, and reconnaissance as well as remote sensing systems. Equipped with commercial-off-the-shelf (COTS) EO/IR and LIDAR sensors, they are capable of providing high resolution 3D mapping of wide geospatial areas and generate a large amount of data. For example, new COTS LIDAR sensors are capable of producing between 20,000 to 200,000 points per second for large range systems and more than a million points per second for short-range systems. The availability of large volumes of information-rich sensor data poses unique challenges to analysts for timely data exploitation. The technologies proposed here would advance the state-of-the-art in the exploitation of LIDAR sensors and will enable robust and efficient detection and classification of objects-of-interest in 3D LIDAR imagery. The proposed technologies will also provide a natural framework for combining context information from other sensors (e.g. EO/IR) for improved target detection and classification. The proposed technologies have wide ranging commercial and defense applications that include:Multi-sensor Intelligence, Surveillance and Reconnaissance Systems.Autonomous surveying systems for urban planningDamage analysis tools for disaster assessment, response and recovery.The proposed technologies advance the state of the art in the DoD S&T emphasis area of AutonomyThe proposed detection and classification technology will result in improved manned and un-manned sensor exploitation systems.