Synthetic Lidar Imaging for Training Robust Deep Learning Classifiers in Degraded Visual Environments
Abstract
Degraded Visual Environment (DVE) encompasses challenging atmospheric or environmental conditions that diminish visibility, posing safety challenges for aircraft and ground vehicles due to factors like dust, fog, smoke, rain, and other obscurants. An effective remote sensor in the DVE is the Light Detection and Ranging, Lidar. However, it often faces challenges in perceiving through DVE. The main goal of this project is to overcome the existing limitations in high-performance commercially available Lidars and present a practical robust solution to mitigate this deficiency. The project has three main goals. Firstly, it explores the concept of wave-particle duality and the interactions between EM waves and matter, particularly how accelerated photonic lights interact with airborne particulates with different physical properties, shape, and sizes. Such interactions lead to light scattering, diffraction, and interference effects and impact Lidar s obscurant and introduce inherent Lidar noise. The primary objective is to conduct Monte Carlo studies to establish statistical noise models representing the Lidar s performance limitations in various Degraded Visual Environment (DVE) conditions. The second project objective is to create a functional laboratory-scale DVE simulator with transparent enclosures. Inside the simulator, powerful air blowers will simulate circular wind forces similar to rotor-wing aircraft. Various particulate matters like sand, fog, and vapor will be introduced inside the simulator to mimic a turbulence DVE state. The simulator will house physical Lidars at one end and target objects at the other. By controlling key measurable process parameters, real-time responses from three high-performance Lidar sensors will be collected and analyzed systematically to establish their noise characteristic models experimentally. Additionally, we will develop efficient voxel modeling algorithms for Lidar 3D point clouds characterization such as terrestrial surface elevation and geometrical irregularity recognition, as well as ground and air obstacles recognition, localization, and characterization. These collected Lidars# responses establish our real-time training datasets for the training of the proposed Deep Learning (DL) classifiers. The third objective is to enhance Lidar s performance through implementation of an active AI-based noise cancellation technique in conjunction with a robust deep learning convolutional neural net classifier for recognition of 3D obstacle objects. To achieve this, we will develop an advanced Generative Adversarial Net, trained using both synthetically and experimentally generated Lidar imagery datasets that encompass various DVE operating conditions. We will utilize PI s developed EM modeling and simulation software, called IRIS-EM, to create a comprehensive virtual testbed for large-scale complex EM scene rendering. IRIS-EM has an ability to simulate functions of different multimodal remote sensors such as SAR, Lidar, Radar, and EO/IR. This feature allows us to systematically generate synthetic multimodal sensor imagery datasets with varying contexts, target objects of interest, and environmental clutter. To ensure high fidelity synthetic Lidar responses from simulator, we will leverage our previously developed and validated noise model(s). This will result in the production of realistic and accurate Lidar imagery, forming the foundation for training the deep learning classifier. The trained classifier will facilitate the creation of transfer learning models, which in turn will streamline the training process of future onboard aircraft classifiers. The ultimate objective is to recruit, mentor, and train minority and underrepresented students to become a skilled workforce for potential employment by the Department of Defense (DoD). Additionally, the research findings will be disseminated through scholarly paper publications and presentations at defense-related conferences and symposiums.
Document Details
- Document Type
- DoD Grant Award
- Publication Date
- Nov 21, 2023
- Source ID
- N000142412006
Entities
People
- Amir Shirkhodaie
Organizations
- Office of Naval Research
- Tennessee State University
- United States Navy