Method for minimizing entropy in hidden Markov models of physical signals

• US 6,212,510 B1
• Filed: 01/30/1998
• Issued: 04/03/2001
• Est. Priority Date: 01/30/1998
• Status: Expired due to Term

• Current Assignee: Hanger Solutions, LLC (IP Investments Group LLC)
• Original Assignee: Mitsubishi Electric Research Laboratories, Inc. (Mitsubishi Electric Corporation)
• Inventors: Brand, Matthew E.
• Primary Examiner(s): Powell, Mark R.
• Assistant Examiner(s): STARKS, WILBERT L
• Application Number:
  US09/016,375
• Time in Patent Office: 1,159 Days
• Field of Search: 706/12, 706/13, 706/15, 706/62
• US Class Current: 706/62
• CPC Class Codes: G06T 2207/10016 Video; Image sequence
• CPC Class Codes: G06T 2207/20076 Probabilistic image processing
• CPC Class Codes: G06T 2207/20081 Training; Learning
• CPC Class Codes: G06T 2207/30232 Surveillance
• CPC Class Codes: G06T 7/277 involving stochastic approaches, e.g. using Kalman filters
• CPC Class Codes: G06V 20/52 Surveillance or monitoring of activities, e.g. for recognising suspicious objects recognising microscopic objects G06V20/69
• CPC Class Codes: G06V 40/23 Recognition of whole body movements, e.g. for sport training

Abstract

A system which observes the world through a video camera and/or other sensors, automatically learns a probabilistic model of normative behavior through the use of a Hidden Markov Model, and uses that model to infer the kind of activity currently under view and detect unusual behavior. The use of hidden Markov models is made possible by entropic training of the model with an θ^θ entropic estimator that folds structure learning into the parameter estimation process to remove parameters from the Hidden Markov Model which have little information content, thus to permit real time robust unusual behavior detection. In one embodiment, the system consists of three components: image analysis; model learning; and signal analysis. In image analysis, each frame of video is reduced to a vector of numbers which describe motion of objects in front of the camera, with a sequence of such vectors, one for each frame of video, establishing the “signal.” In model learning, the signal is analyzed to obtain parameters for a probabilistic model of the dynamics of the scene in front of the camera. In signal analysis, the model is used to classify and/or detect anomalies in signals produced on-the-fly by image analysis of new video.

Claims

1. A computer implemented method for modeling a behavior of a signal, comprising the steps of:

reducing a training signal to a sequence of vectors;

storing the sequence of vectors in a memory as a hidden Markov model including states and state transitions;

providing an entropic prior probability; and

estimating maximum a posteriori probability parameters of the stored hidden Markov model using the entropic prior probability to retain states and state transitions indicative of a normative behavior of the training signal resulting in a low entropy hidden Markov model stored in the memory.

2. The method of claim 1 wherein the entropic prior probability has a minimum at zero and a maximum, and further comprising the steps of:

gradually forcing the entropic prior probability to either zero or one; and

discarding states and state transitions having a probability substantially close to zero.

3. The method of claim 1 further comprising:

providing an observation signal;

evaluating the observation signal with the low entropy hidden Markov model;

generating an alarm signal when the low entropy hidden Markov model evaluates the observation signal as having substantially low probabilities.

4. The method of claim 3 wherein the training signal and the observation signal each is a video sequence including a plurality of frames.

5. The method of claim 1 wherein the estimating step further comprises the step of:

setting a derivative of a log-posterior probability to zero, the log-posterior probability being derived from a likelihood function of the hidden Markov model and the entropic prior probability.

6. The method of claim 1 wherein the training signal is a video sequence of a scene, and the video includes a plurality of frames, and the vectors are motion vectors of pixel movement in the frames, and the normative behavior is a normative behavior in the scene.

7. The method of claim 6 further comprising the steps of:

determining first and second order moments of the pixel movement; and

determining histograms of the pixel movement.