Deep Learning for Physics Based Models of 3D Images in Degraded Environment: Algorithms, Mathematica
Abstract
Degraded environments such as fog, brownout conditions, extremely low light, counter measures, camouflage, and occlusions present su,bstantial difficulties in scene understanding and automated object recognition from visual data. Deep learning is a technological re,volution applied successfully in many disciplines including automated object recognition. However, most deep learning models are dev,eloped for visual data without the presence of substantial degradations as described earlier, particularly for multiple dimensional,(3D + time domain) data. In addition, many researchers have concentrated on developing computational algorithms for understanding su,rveillance imagery and/or object recognition without considering the physical model of visual data from visual sources in the presen,ce of severe degraded environment. Understanding and developing the physics based models of degraded environments for visual sources, are important for data augmentation in deep learning to substantially enhance the performance of such systems.In this proposal, we,propose mathematical models based on the physics of the complex multidimensional scenery in the presence of degraded environmental c,onditions to be used in conjunction with deep learning. A variety of 3D sensing models for scene capture will be considered includin,g passive 3D image capture, and LiDAR (RGB+Depth). We aim to investigate this approach and assess its success in the presence of the, environmental degradations for applications including scene understanding, visualization and recognition, and in particular gesture, and activity recognition which are important blocks for inferring intentions. There are no extensive or even limited database for 3,D data from visual sources in low light, 3D data in fog, counter measures, camouflage, or 3D data in environmental degradations in g,eneral. Thus, our additional plans are to develop accurate physical model of 3D plus time domain data in degradation environment whi,ch are also critical in effective implementation of deep learning systems for the problems of interest to the NAVY warfighter. Overc,oming these adversities in degraded environments have substantial benefits to the NAVY warfighter and positive impact on DoD capabil,ities in broad applications including scene understanding from visual data, object recognition, human activity recognition, event re,cognition, and inferring hostile intensions. Approved for Public Release
Document Details
- Document Type
- DoD Grant Award
- Publication Date
- May 16, 2022
- Source ID
- N000142212349
Entities
People
- Bahram Javidi
Organizations
- Office of Naval Research
- United States Navy
- University of Connecticut