Robustness Studies in 3D Camera Data
Abstract
3D cameras such as RGB-D (color + depth) and LiDAR (Light Detection and Ranging) cameras have enabled numerous computer vision applications such as crime scene reconstruction, damage detection, surveillance, autonomous vehicles etc. Tampering with data from 3D cameras may result in improper functioning of these applications. Most existing research, however, have focused on developing new approaches that address problems such as real-time requirements and high quality 3D reconstruction. There does not seem to be much work done to detect possibility of tampering and authenticate the data streams from 3D cameras. The main aim of this proposal is to create a framework that can test the robustness of the data from 3D cameras. We propose a transformative approach, Robust 3D (R3D), that uses an anti-forensic 3D object stream manipulation framework to create attacks on data generated by 3D cameras and then build a forensic system that can authenticate the validity of the generated data in the presence of anti-forensic attacks. Towards this end, we have carried out few preliminary works: ¥ Anti-forensics framework that uses computer vision and graphics methods to tamper the data generated from 3D LiDAR and RGB-D cameras. ¥ Experiments conducted with a set of users (who are computer vision and graphics scientists) carrying out a visual inspection of the manipulated RGB-D streams (just like security personnel would do). ¥ Efficacy of forensic approaches on the manipulated 3D camera data. These preliminary studies show that it was very difficult to distinguish between the real or reconstructed rendering of such 3D video sequences, thus clearly showing the potential risk involved. The preliminary forensic techniques handle simple attacks, however, they are still vulnerable to more sophisticated attacks. Research Plan: We propose an iterative development of R3D framework creating more sophisticated attacks to tamper the 3D camera data and improving the forensic techniques to detect such attacks. For RGB-D camera data, we will develop an anti-forensic method, DISPOSE (DIstortion Score based POse mesh SElection), for minimizing the artifacts that arise due to the manipulations in the 3D scenes. This method uses a database of 3D meshes to select the mesh that would be more appropriate for the desired output animation with the highest possible quality. For LiDAR data, we will develop approaches that will help us carry out additive, subtractive, and de-formative attacks. We will then develop real-time forensic techniques that can detect such tampering. The manipulated 3D camera data streams will then be subjected to various kinds of forensic analysis and user evaluation. We hypothesize that by carrying out such an iterative set of anti-forensic and forensic analysis, we will be able to create a robust framework for applications using 3D camera data, with the following scientific merits: ¥ Real-time distortion score based 3D pose mesh selection approach to generate manipulated RGB-D camera data. ¥ Real-time framework for tampering 3D LiDAR data using additive, subtractive, and de-formative attacks. ¥ Real-time forensic algorithms for detecting possible tampering of 3D scenes generated from RGB-D and LiDAR cameras and authenticate the data if there are no evidences of tampering. Apart from sharing the research results through refereed conference and journal publications, the algorithms developed will be made available as anti-forensics and forensics tools for appropriate researchers. The PI teaches courses on Multimedia Systems and Video Analytics. Results and experiences obtained through this research will be incorporated into the curriculum of these courses.
Document Details
- Document Type
- DoD Grant Award
- Publication Date
- Sep 11, 2018
- Source ID
- W911NF1710299
Entities
People
- Balakrishnan Prabhakaran
Organizations
- Army Contracting Command
- United States Army
- University of Texas at Dallas