|
Scene Modeling for Far-Field Activity Analysis
Tomas Izo & Eric Grimson
Fig. 1: Example scene model--a sample of ~3000 moving object tracks
(left) grouped into ~100 clusters (right). Tracks with different color
belong to different clusters (colors are reused).
 |
Introduction
A key goal of any visual surveillance system is to automatically determine
when an observed scene contains unusual or unexpected activity. In the
past this task was performed by a human expert: someone familiar with
the scene who was able to recognize when something out of the ordinary
occurred. A typical surveillance site may have so many sensors in different
locations that it is no longer feasible for a person to monitor all of
them. Machine vision systems are needed to mine the collected data for
potentially interesting activity. This has fostered a new area of machine
vision research, aimed at building statistical models of the usual pattern
of activity in scenes.
Methods
We define an activity as the movement of a single object, such as a car,
person or group of people, through the scene. Various aspects of activities,
to which we refer as tracks, include the paths the objects take, their
sizes, velocities, direction of motion, etc. We propose a method for unsupervised,
multi-featured learning of a statistical model of activities that can
make use of a very large data set for learning, yet enables fast reasoning
about new tracks, both partial and complete. We use a versatile similarity
measure that groups together tracks only when they are similar along all
of the various observed aspects. We group tracks using spectral clustering
and estimate the spectral embedding efficiently from a sample of tracks
using the Nystrom approximation [1]. Clusters are modeled as Gaussians
in the embedding space and new tracks are projected into the embedding
space and matched with the cluster models to detect anomalies. The ability
to reason about partial tracks makes it possible to detect surprising
moments, which occur when there is a sudden change in the belief distribution
for a given track.
Results
To learn a statistical model of activity in the scene, we use a set of
approximately 40,000 moving object tracks, corresponding to an entire
week of activity in a busy urban outdoor scene. Figure 2 shows two examples
of unusual activities detected using the learned model by thresholding
on the likelihood of each activity under the model. In addition to obtaining
qualitative results such as the ones in Figure 2, we also demonstrate
the validity of our scene modeling and anomaly detection framework by
collecting human judgments for a sample of several hundred examples from
the data set and showing that they are correlated with the likelihoods
of the examples under the model.
Fig. 2: Examples of detected unusual activity. Left: A person walking
along an unusual path and dropping off a large object (suddenly changing
size). Right: A car driving into the pedestrian zone and turning around.
 |
Research Support
This project is supported in part by a grant from DARPA.
References:
[1] Charles F. Fowlkes, Serge Belongie, Fan Chung and Jitendra Malik.
Spectral grouping using the Nystrom method. In IEEE Trans. Pattern
Analysis and Machine Intelligence, 26(2):214-225, 2004.
|

 |