Learning Models for Predictive Behavioral Intent and Activity Analysis in Wide Area Video Surveillance

• NCJ Number: 250273
• Author(s): Shishir K. Shah
• Date Published: September 2015
• Length: 16 pages
• Grant Number(s): 2009-MU-MU-K004
• Agencies: NIJ-Sponsored
• Publication Type: Research (Applied/Empirical); Report (Study/Research); Report (Grant Sponsored); Grants and Funding
• Award: Learning Models for Predictive Behavioral Intent and Activity Analysis in Wide Area Video Surveillance

Annotation

Methodology and outcomes are reported for a research project that focused on the development of new methods and algorithms for wide-area surveillance and understanding of human activities in large distributed camera networks.

Abstract

The project's main product was research publications that span algorithms for tracking individual and group activity and the re-identification of individuals and groups. Each of the algorithms was developed and tested on public and small benchmark datasets. These are stand-alone evaluations that have not yet been integrated into a system solution. A prototype code was developed for person re-identification that facilitates production of a gallery probe from input videos and performs matching of an input observation to the identifications in the gallery. Although this is not an integrated system for re-identification, the developed prototype has two modular components. One module takes in a video and performs human detection and tracking. Tracking allows for the production of either a gallery or probe dataset. The second module includes human parts-based re-identification. The images in the specified gallery and probe datasets are initially segmented to identify body parts. An appearance descriptor is produced as a model for each body part and integrated over multiple images of the person, if available. This technology provides automatic monitoring of surveillance videos, which helps security officers detect and predict suspicious activities; this facilitates the prevention or mitigation of a security threat. The ability to model simple activity patterns from a single camera's video data has been demonstrated for a variety of applications. The challenges that remain are related to person re-identification and the understanding of human behavior. The latter is related to an overall lack of understanding of human behavior, both as individuals and groups. 4 figures and 4 references