Opportunistic cascade and cascade training, evaluation, and execution for vision-based object detection

Patent metadata

Field	Value
Publication number	US-9449259-B1
Application number	US-201213558298-A
Country	US
Kind code	B1
Filing date	Jul 25, 2012
Priority date	Jul 25, 2012
Publication date	Sep 20, 2016
Grant date	Sep 20, 2016

Abstract

The present invention relates to a classifier cascade object detection system. The system operates by inputting an image patch into parallel feature generation modules, each of the feature generation modules operable for extracting features from the image patch. The features are provided to an opportunistic classifier cascade, the opportunistic classifier cascade having a series of classifier stages. The opportunistic classifier cascade is executed by progressively evaluating, in each classifier in the classifier cascade, the features to produce a response, with each response progressively utilized by a decision function to generate a stage response for each classifier stage. If each stage response exceeds a stage threshold then the image patch is classified as a target object, and if the stage response from any of the decision functions does not exceed the stage threshold, then the image patch is classified as a non-target object.

First claim

What is claimed is:

1. A classifier cascade object detection system, comprising a non transitory memory and one or more processors, the non transitory memory having executable instructions encoded thereon, such that upon execution of the instructions by the one or more processors, the one or more processors perform operations of:
inputting an image patch into parallel feature generation modules, each of the feature generation modules operable for extracting different features from the image patch;
providing the different features to a classifier cascade, the classifier cascade having a series of different classifiers; and
executing the classifier cascade by progressively evaluating, in each classifier in the classifier cascade, one or more of the different features to produce a response, if each classifier produces a response that exceeds a predefined threshold then the image patch is classified as a target object, and if the response from any of the classifiers in the classifier cascade does not exceed the predefined threshold, then the image patch is classified as a non-target object;

wherein the classifier cascade is trained by performing operations of:
loading parameters of the image patch, including window position (x,y) with respect to a larger image and window height (h);
loading a cascade template file, the cascade template file defining a number of classifier stages in the classifier cascade, one or more feature types, a classifier type to use in each classifier stage, an aspect ratio of the image patch, and a height of an image storage container which all image patches are re-sampled to;
computing features based on features specified in the cascade template file;
compiling the features for a classifier trainer, the classifier trainer generating a trained classifier;
generating a Receiver-Operating-Characteristics (ROC) curve, which includes false-alarm rate and true-detection rate pairs (FAR,TDR) for a given stage threshold; and
tuning the stage threshold.

2. The classifier cascade object detection system as set forth in claim 1 , wherein the classifier is an opportunistic classifier cascade, such that the opportunistic classifier cascade has a series of classifier stages; and
wherein executing the classifier cascade is further performed by executing the opportunistic classifier cascade by progressively evaluating, in each classifier in the classifier cascade, the features to produce a response, with each response progressively utilized by a decision function to generate a stage response for each classifier stage, such that if each stage response exceeds a stage threshold then the image patch is classified as a target object, and if the stage response from any of the decision functions does not exceed the stage threshold, then the image patch is classified as a non-target object.

3. The classifier cascade object detection system as set forth in claim 2 , wherein the decision function utilizes a weight sum of the classifier responses of a current classifier stage and previous classifier stages, as follows:

f
n

⁡

(

h
0

,

h
1

,
…

)

=

∑

i
=
0

n

⁢

α

n
,
i

·

h
i

,

where f n is the stage response for stage n, α n is a weight, and h i is a classifier response.

4. The classifier cascade object detection system as set forth in claim 3 , wherein if f n >τ n , where τ n is a stage threshold for stage n, the image patch is analyzed by a next stage or classified as a target object; otherwise, the image patch is classified as a non-target object.

5. The classifier cascade object detection system as set forth in claim 4 , wherein weights an are set to all ones.

6. The classifier cascade object detection system as set forth in claim 5 , wherein the stage threshold is tuned by finding the threshold τ* required to achieve the specified target stage true-detection rate (Target TDR), using the ROC curve.

7. The classifier cascade object detection system as set forth in claim 6 , wherein the stage threshold is tuned by performing operations of:
setting a stage threshold is set such that:
τ*={τ: TDR=Target TDR}

where TDR is the True Detection Rate curve function for a current classifier stage, such that the target TDR is a parameter used to tune a number of examples to be further analyzed by subsequent classifier stages.

8. The classifier cascade object detection system as set forth in claim 1 , further comprising a feature vector recall mechanism, such that features that have been computed in earlier stages in the classifier cascade are made available to other succeeding stages in the classifier cascade that use the same feature.

9. The classifier cascade object detection system as set forth in claim 1 , wherein each classifier in the classifier cascade receives an input that is distinct from inputs received by other classifiers in the classifier cascade.

10. A computer implemented method for object detection using a classifier cascade, comprising an act of causing a processor to execute instructions encoded upon a non transitory memory, such that upon execution, the processor performs operations of:
inputting an image patch into parallel feature generation modules, each of the feature generation modules operable for extracting different features from the image patch;
providing the different features to a classifier cascade, the classifier cascade having a series of different classifiers; and
executing the classifier cascade by progressively evaluating, in each classifier in the classifier cascade, one or more of the different features to produce a response, if each classifier produces a response that exceeds a predefined threshold then the image patch is classified as a target object, and if the response from any of the classifiers in the classifier cascade does not exceed the predefined threshold, then the image patch is classified as a non-target object;

wherein the classifier cascade is trained by performing operations of:
loading parameters of the image patch, including window position (x,y) with respect to a larger image and window height (h);
loading a cascade template file, the cascade template file defining a number of classifier stages in the classifier cascade, one or more feature types, a classifier type to use in each classifier stage, an aspect ratio of the image patch, and a height of an image storage container which all image patches are re-sampled to;
computing features based on features specif

Assignees

• Cheng Shinko Y
• Owechko Yuri
• Medasani Swarup
• Hrl Lab Llc

Inventors

• Cheng Shinko Y
• Owechko Yuri
• Medasani Swarup

Classifications

• G06V10/7747

  Organisation of the process, e.g. bagging or boosting · CPC title

• G06F18/2148

  characterised by the process organisation or structure, e.g. boosting cascade · CPC title

• G06N20/00 Primary

  Machine learning · CPC title

• G06F18/21

  Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation · CPC title

• G06V10/255

  Detecting or recognising potential candidate objects based on visual cues, e.g. shapes · CPC title