Explainable Multi-Source Fusion in Deep Learning for Explosive Hazard Detection
Abstract
Major Goals: Explosive hazards (EH) pose severe threats as remnants of past and current conflict zones, endangering the safety of both civilian and military personnel. The U.S. Army is actively engaged in the exploration of innovative methods for the detection and removal of these hazardous devices. Over the last few decades, our groups focus has encompassed various approaches, including hand-held and ground vehicle-based deployment, such as forward- and side-looking EH detection (EHD). Our previous efforts and objectives are detailed in the initial ARO proposal. Throughout the duration of this grant, we collaborated closely with the Army DEVCOM RTI, at Fort Belvoir, to delve into specific areas pertinent to a novel EH deployment platform -- using drones equipped with multi-sensor payloads. The list of topics from our initial proposal include: 1. Detection and localization - Automatic identification of EH threats from aerial sensors. 2. Scene understanding - Incorporation of platform meta-data and environmental meta-data for context to drive advanced EHD reasoning from an aerial platform. 3. Multi-sensor fusion - Machine learning and data fusion in and across aerial sensors (RGB, infrared, multispectral, polarized cameras, etc.). 4. Data augmentation - EHD is generally plagued by limited data and a need to operate in a range of environments. In return, we are focused on algorithms that can learn from class imbalanced small data sets and techniques to artificially create training data. 5. Explainability - The above theories need to be understandable. Black box solutions cannot be trusted nor expected to generalize. In the remainder of this report, we summarize our overall efforts. Specifically, we discuss progress in terms of open published academic research. If ARO would like additional details, they can be found in our biweekly presentations. Matt Aeillo and Clare Yang at RTI are our technical points of contact (TPOC).
Document Details
- Document Type
- Technical Report
- Publication Date
- Aug 14, 2023
- Accession Number
- AD1230002
Entities
People
- Derek T. Anderson
Organizations
- University of Missouri