METHOD FOR AUTOMATIC DETECTION AND TRACKING OF MULTIPLE TARGETS WITH MULTIPLE CAMERAS AND SYSTEM THEREFOR

• US 20100166260A1
• Filed: 06/10/2009
• Published: 07/01/2010
• Est. Priority Date: 12/25/2008
• Status: Active Grant

• Current Assignee: National Chiao Tung University (Government of The Republic of China)
• Original Assignee: National Chiao Tung University (Government of The Republic of China)
• Inventors: HUANG, Ching-Chun, Wang, Sheng-Jyh
• Granted Patent: US 8,995,712 B2
• Time in Patent Office: Days
• US Class Current: 382/103
• CPC Class Codes: G06V 10/22 by selection of a specific region containing or referencing a pattern; Locating or processing of specific regions to guide the detection or recognition
• CPC Class Codes: G06V 10/255 Detecting or recognising potential candidate objects based on visual cues, e.g. shapes
• CPC Class Codes: G06V 20/52 Surveillance or monitoring of activities, e.g. for recognising suspicious objects recognising microscopic objects G06V20/69

Abstract

A method for automatically detecting and tracking multiple targets in a multi-camera surveillance zone and system thereof. In each camera view of the system only a simple object detection algorithm is needed. The detection results from multiple cameras are fused into a posterior distribution, named TDP, based on the Bayesian rule. This TDP distribution represents a likelihood of presence of some moving targets on the ground plane. To properly handle the tracking of multiple moving targets with time, a sample-based framework which combines Markov Chain Monte carlo (MCMC), Sequential Monte Carlo (SMC), and Mean-Shift Clustering, is provided. The detection and tracking accuracy is evaluated by both synthesized videos and real videos. The experimental results show that this method and system can accurately track a varying number of targets.

Claims

1. A method of multi-target detecting and tracking with multiple cameras, comprising the steps of:

(a) analyzing a detection result of a plurality of camera views using a target detection algorithm to generate a plurality of analysis data;

(b) integrating the analysis data with a Baysian framework to create a target detection probability (TDP) distribution; and

(c) simultaneously and automatically detecting and tracking a plurality of moving targets in the camera views.

2. The method as claimed in claim 1, wherein the camera views are 2D camera views and the analysis data are 2D analysis data.

3. The method as claimed in claim 1, wherein the TDP distribution is a posterior probability distribution representing a likelihood of a presence of each moving target at a ground location in accordance with a set of foreground images from a plurality of cameras.

4. The method as claimed in claim 1, wherein the target detection algorithm in the step (a) is applied to a 2D background subtraction between any one of the camera views and a background thereof.

5. The method as claimed in claim 1, wherein the TDP distribution in the step (b) equals to a sum of a probability, G₁(X), making a newly appearing target to be detected easily and a probability, G₂(X), providing temporal information in the time domain between successive frames, and X is a location on a ground plane.

6. The method as claimed in claim 1, wherein the step (c) is a sample management process, further comprising:

(c-1) a sample generating step of executing a MCMC sampling corresponding to the probability G₁(X) and generating a plurality of samples from the TDP distribution, and executing a SMC sampling corresponding to the probability G₂(X) to iteratively update and predict each respective location, weight and ID of the samples;

(c-2) a sample labeling step of adopting a maximum likelihood decision rule to classify the samples;

(c-3) a new target identifying step of grouping similar samples with a mean shift clustering method to label a new target with a new ID and verify the new target; and

(c-4) a target updating step of assigning a unique ID to each new sample, adding 1 to the number of the targets, or when the number of the samples for targets is lower than a preset threshold, removing the target and subtracting 1 from the number of the targets.

7. A system of multi-target detecting and tracking with multiple cameras, comprising:

a detection analyzing means for analyzing a plurality of camera views using a target detection algorithm to generate a plurality of analysis data;

a Bayesian framework for integrating the analysis data to form a TDP distribution so as to detect any newly appearing target and provide temporal information in a time domain between successive frames; and

a tracking means for automatically detecting and tracking a plurality of moving targets in the camera views and converting the detecting and tracking into a sample management process.

8. The system as claimed in claim 7, wherein the camera views are 2D camera views, and the analysis data are 2D analysis data.

9. The system as claimed in claim 7, wherein the TDP distribution expresses is a posterior probability distribution representing the a likelihood of a presence of each moving target at a ground location in accordance with a set of foreground images from a plurality of cameras.

10. The system as claimed in claim 7, wherein the target detection algorithm is applied to a 2D background subtraction between any one of the camera views and a background thereof.

11. The system as claimed in claim 7, wherein the TDP distribution equals to a sum of a probability, G₁(X), making a newly appearing target to be detected easily and a probability, G₂(X), providing temporal information in the time domain between successive frames, and X is a location on a ground plane.

12. The system as claimed in claim 11, wherein the tracking means comprises:

a sample generating module for executing a MCMC sampling corresponding to the probability G₁(X) to generate a plurality of samples from the TDP distribution, and for executing a SMC sampling corresponding to the probability G₂(X) to iteratively update and predict each respective location, weight and ID of the samples;

a sample labeling module for adopting a maximum likelihood decision rule to classify the samples;

a new target identifying module for grouping similar samples with using a mean shift clustering method to label a new target with a new ID and verify the new target; and

a target updating module for assigning a unique ID to any new sample, adding 1 to the number of target, or when a number of sample for a target is lower than a preset threshold, removing the target and subtracting 1 from the number of target.