Find and Track People in Real Video Imagery

We are engaged in designing, implementing, and testing a system that can detect and track humans automatically. Our system will recognize the activities of individuals and patterns of activities within and between groups. This information could be used to provide alerts of potential threats to facilities and personnel.
Successfully representing human activities requires a representation of human motion at the kinematic level. This proposal requests funds to support an effort that will refine our tracking algorithms and improve the kinematic expresentations that they produce. In particular, we request support for our efforts to build reliable, self-starting kinematic trackers. We will build self-initializing kinematic trackers that use known coherence in the structure and movement of people to detect people and track them. Our process involves a series of steps going from coarse to fine:
Finding segments. Human body segments are identified by the fact that they are coherent in color, texture and motion; that they have a predictable shape; and that they appear in a series of images.Forming kinematic assemblies. Segments in each frame are assembled into groups that could be a view of a person.
Exploiting motion coherence. Possible tracks (i.e., those that could be people) are constructed from segments that move coherently from frame to frame and form assemblies in multiple frames; these assemblies must have reasonable frame-frame motions. Kinematic and dynamic refinement. Tracks are fed to a kinematic tracker that refines the estimates of configuration and compares these detailed estimates of motion with possible human movements.