Explainable Multi-Source Fusion in Deep Learning for Explosive Hazard Detection
Abstract
This proposal is in response c to section c.i.2, (Information Processing & Fusion) in the U.S. Army BAA W911NF-17-S-002(TPOC is Dr. Liyi Dai). The aim of this research is to significantly advance the U.S. Army s understanding of and ability to detect explosive hazards (landmines, IEDs, and UXOs) via machine learning (ML) and multi-source (so sensor, human and algorithm) data/information fusion. Buried explosives remain a major threat to both soldiers and civilians in regions targeted by terrorist and criminal organizations because buried explosives are relatively cheap and easy to make using commonly available materials. While current technologies are a success, their positive detection (PD) and false alarm rates (FAR) need significant improvement. This work will explore hand held (HH), forward looking (FL), and side looking (SL) platforms. Sensing modalities include, but are not limited to, acoustic, infrared (IR), ground penetrating radar (GPR), voxel space radar, electromagnetic induction (EMI), and precise positioning. This research is supported by the Night Vision and Electronic Sensors Directorate (NVESD), in terms of sensors, platforms, data collection and discussions regarding project findings if/when appropriate. The goal is to explore novel ML and fusion mathematics (and associated algorithms) to make needed leaps in explosive hazard detection (EHD). Technical objectives include: 1. Data fusion in deep learning (DL): discover innovative ways to conduct explainable fusion in and across DLs at different levels, e.g., signal, feature and decision. 2. DL on limited (volume and variety) data that has a class imbalance: explore architectures, training, transfer learning and generative adversarial networks (GANs) for sensor fusion. 3. Multi-sensor fusion: study the application of 1) and 2) for fusing sensors on various platforms and use those results to improve 1) and 2) (the underlying math and algorithms) and unearth new domain knowledge (technology transfer centered on the sensors, platforms, targets, context, etc.). 4. Experimentation: extensive continuous performance testing with respect to a wide range of sensors, target, burial depths, soils, environments, and clutter. The above research is anticipated to lead to a significant improvement in positive detection (PD) and/or false alarm (FA) reduction for EHD. It will give rise to new mathematics for multi-source fusion, signal/image processing and ML. Beyond EHD, the theories put forth and results found will be of interest to multiple Army, DoD and academic technological thrusts; e.g., Big Data, unmanned aerial systems, computer vision and geospatial intelligence, to name a few.
Document Details
- Document Type
- DoD Grant Award
- Publication Date
- Feb 14, 2019
- Source ID
- W911NF1810153
Entities
People
- Derek T. Anderson
Organizations
- Army Contracting Command
- United States Army
- University of Missouri