Multimodal Signal Processing for Personnel Detection and Activity Classification for Indoor Surveillance
Abstract
This goal of this project was to develop novel schemes for the fusion of heterogeneous information. The target application was the detection and classification of personnel activity in both indoor and outdoor environments under dependent observations. We have identified features and designed a classifier that achieves up to 95% classification accuracy on classifying the occupancy with indoor footstep data. MDL-based copula selection strategies are investigated and a detector based on vines is designed that extends previous bivariate copula-based detectors to a multi-sensor application. Our copula-based detectors yield more than 40% improvement over the conventional data fusion methodologies. We extend our solution to quantized sensor information and demonstrate that injecting controlled noise can dramatically reduce computational complexity with insignificant performance loss. We propose and derive the Conditional Posterior Cram r-Rao Lower Bound (CPCRLB) for online tracking. We demonstrate that the PCRLB-based iterative approach converges quickly with significantly reduced computational cost as compared to a one-shot approach. Detector design in the presence of security threats, such as data falsification attacks to sensor networks, are also addressed. Error-control codes and decoding algorithms are used to reliably classify data in a network containing human agents.
Document Details
- Document Type
- Technical Report
- Publication Date
- Nov 15, 2013
- Accession Number
- ADA606602
Entities
People
- Pramod Varshney
Organizations
- Syracuse University