Method and apparatus for detecting generalized passerby by utilizing features of a wheel and upper body

• US 9,519,831 B2
• Filed: 12/08/2014
• Issued: 12/13/2016
• Est. Priority Date: 12/27/2013
• Status: Active Grant

• Current Assignee: Neusoft Reach Automotive Technology (Shanghai) Co., Ltd.
• Original Assignee: Neusoft Corporation
• Inventors: Chai, Liying, Yu, Bing, Liu, Wei, Duan, Chengwei, Yuan, Huai
• Primary Examiner(s): Heidemann, Jason
• Assistant Examiner(s): Shin, Brian
• Application Number:
  US14/562,771
• Publication Number:
  US 20150186723A1
• Time in Patent Office: 736 Days
• Field of Search: None
• US Class Current: 1/1
• CPC Class Codes: G06F 18/24 Classification techniques
• CPC Class Codes: G06V 20/58 Recognition of moving objects or obstacles, e.g. vehicles or pedestrians; Recognition of traffic objects, e.g. traffic signs, traffic lights or roads
• CPC Class Codes: G06V 40/10 Human or animal bodies, e.g. vehicle occupants or pedestrians; Body parts, e.g. hands

Abstract

A method for detecting a generalized passerby includes: acquiring an input image; determining whether a preset common feature of a wheel exists in the input image; selecting an image window at left side or right side or upper side of a center of a region where the preset common feature of the wheel is located in a case that the preset common feature of the wheel exists in the input image; inputting the selected image window into a preset upper-body classifier; detecting whether an upper body of a passerby exists in the selected image window and outputting a first detection result.

Claims

1. A method for detecting a generalized passerby, comprising:

acquiring an input image;

determining whether a preset common feature of a wheel exists in the input image;

selecting an image window at left side or right side or upper side of a center of a region where the preset common feature of the wheel is located in a case that the preset common feature of the wheel exists in the input image; and

inputting the selected image window into a preset upper-body classifier, detecting whether an upper body of a passerby exists in the selected image window and outputting a first detection result,

wherein the preset upper-body classifier is generated by training with a preset common feature of a generalized passerby, the preset common feature of the generalized passerby comprises a common feature of a human body, and the preset common feature of the wheel comprises a circle feature and an oval feature,

wherein selecting the image window at left side or right side or upper side of the center of the region where the preset common feature of the wheel is located comprises;

determining a size of the image window based on a radius of the region where the preset common feature of the wheel is located; and

determining a position of the image window based on the radius of the region where the preset common feature of the wheel is located and the center of the region where the preset common feature of the wheel is located.

2. The method for detecting the generalized passerby according to claim 1, further comprising:

pre-processing the input image, inputting the pre-processed image into a preset whole-body classifier, detecting whether a whole body of a passerby exists in the pre-processed image and outputting a second detection result, wherein the preset whole-body classifier is generated by training with the preset common feature of the generalized passerby and the preset common feature of the generalized passerby comprises the common feature of the human body.

3. The method for detecting the generalized passerby according to claim 2, wherein pre-processing the input image and inputting the pre-processed image into the preset whole-body classifier comprises:

acquiring at least two images having different resolutions by processing the input image using a pyramid down-sampling technology, respectively dividing the images having different resolutions into a plurality of small-window images having a same size as those of samples for training the preset whole-body classifier, and inputting the small-window images into the preset whole-body classifier in sequence.

4. The method for detecting the generalized passerby according to claim 1, wherein determining the size of the image window based on the radius of the region where the preset common feature of the wheel is located comprises:

selecting an image window having a width of w=2kr and a length of l=3kr; and

determining the position of the image window based on the radius of the region where the preset common feature of the wheel is located and the center of the region where the preset common feature of the wheel is located comprises;

determining the position of the image window, with the position of the image window being within a value range of [x−

6r, x+6r] in a horizontal direction and within a value range of [y−

3r, y+3r] in a vertical direction,

wherein k is a scaling factor, r is a radius of the region where the preset common feature of the wheel is located, and x and y are coordinates of the center of the region where the preset common feature of the wheel is located.

5. The method for detecting the generalized passerby according to claim 1, wherein the common feature of the human body comprises a head feature, a shoulder feature and a torso feature, wherein

the head feature comprises a head region depicted with a first color and a background region depicted with a second color, each of the head region and the background region is rectangular, the head region is located within the background region and a lower edge of the head region and a lower edge of the background region are coincided, the region depicted with the second color has an area which is N1 times as great as that of the region depicted with the first color, a feature value of the head feature is a1=|b1−

c1|, wherein b1 is a sum of values of pixels in the region depicted with the second color, c1 is N1 times as great as a sum of values of pixels in the region depicted with the first color, and N1 is a positive value;

the shoulder feature comprises a shoulder region depicted with the first color and a background region depicted with the second color, each of the shoulder region and the background region is rectangular, the shoulder region is located at a lower left corner of the background region with a left edge and a lower edge of the shoulder region and a left edge and a lower edge of the background region being coincided respectively, or the shoulder region is located at a lower right corner of the background region with a right edge and a lower edge of the shoulder region and a right edge and a lower edge of the background region being coincided respectively, the region depicted with the second color has an area which is N2 times as great as that of the region depicted with the first color, a feature value of the shoulder feature is a2=|b2−

c2|, wherein b2 is a sum of values of pixels in the region depicted with the second color, c2 is N2 times as great as a sum of values of pixels in the region depicted with the first color, and N2 is a positive value;

the torso feature comprises a vertical torso feature and an inclined torso feature,

the vertical torso feature comprises a first vertical torso feature and a second vertical torso feature, the first vertical torso feature comprises a first vertical torso region depicted with the first color and a background region depicted with the second color, the first vertical torso region and the background region are located in a same rectangle, the region depicted with the second color has an area which is N3 times as great as that of the region depicted with the first color, a feature value of the first vertical torso feature is a3=|b3−

c3|, wherein b3 is a sum of values of pixels in the region depicted with the second color, c3 is N3 times as great as a sum of values of pixels in the region depicted with the first color, and N3 is a positive value;

the second vertical torso feature comprises a second vertical torso region depicted with the first color and a background region depicted with the second color, each of the second vertical torso region and the background region is rectangular, the second vertical torso region is parallel with the background region in a vertical direction, the region depicted with the first color has an area equal to that of the region depicted with the second color, the region depicted with the first color is separated from the region depicted with the second color by M1 pixels, a feature value of the first vertical torso feature is a4=|b4−

c4|, wherein b4 is a sum of values of pixels in the region depicted with the second color, and c4 is a sum of values of pixels in the region depicted with the first color; and

the inclined torso feature comprises an inclined torso region depicted with the first color and a background region depicted with the second color, each of the inclined torso region and the background region is rectangular, an inclination angle of the inclined torso region is equal to an inclination angle of the background region, and the background region is located at left side or right side or upper side of the inclined torso region, the region depicted with the first color has an area equal to that of the region depicted with the second color, the region depicted with the first color is separated from the region depicted with the second color by M2 pixels, a feature value of the inclined torso feature is a5=|b5−

c5|, wherein b5 is a sum of values of pixels in the region depicted with the second color, and c5 is a sum of values of pixels in the region depicted with the first color.

6. An apparatus for detecting a generalized passerby, comprising a processor and a non-transitory processor-readable medium having processor-executable instructions stored thereon that, when executed by the processor, cause the apparatus to:

acquire an input image; and

determine whether a preset common feature of a wheel exists in the input image, select an image window at left side or right side or upper side of a center of a region where the preset common feature of the wheel is located in a case that the preset common feature of the wheel exists in the input image, input the selected image window into a preset upper-body classifier, detect whether an upper body of a passerby exists in the selected image window and output a first detection result, wherein the preset upper-body classifier is generated by training with a preset common feature of a generalized passerby, the preset common feature of the generalized passerby comprises a common feature of a human body, and the preset common feature of the wheel comprises a circle feature and an oval feature,

wherein selecting the image window at left side or right side or upper side of the center of the region where the preset common feature of the wheel is located in a case that the preset common feature of the wheel exists in the input image comprises;

determining a size of the image window based on a radius of the region where the preset common feature of the wheel is located; and

determining a position of the image window based on the radius of the region where the preset common feature of the wheel is located and the center of the region where the preset common feature of the wheel is located.

7. The apparatus for detecting the generalized passerby according to claim 6, wherein the instructions further cause the apparatus to pre-process the input image, input the pre-processed image into a preset whole-body classifier, detect whether a whole body of a passerby exists in the pre-processed image and output a second detection result, wherein the preset whole-body classifier is generated by training with a preset common feature of a generalized passerby, and the preset common feature of the generalized passerby comprises a common feature of a human body.

8. The apparatus for detecting the generalized passerby according to claim 7, wherein pre-processing the input image and inputting the pre-processed image into the preset whole-body classifier comprises acquiring at least two images having different resolutions by processing the input image using a pyramid down-sampling technology, respectively dividing the images having different resolutions into a plurality of small-window images having a same size as those of samples for training the preset whole-body classifier, and inputting the small-window images into the preset whole-body classifier in sequence.

9. The apparatus for detecting the generalized passerby according to claim 6, wherein

determining the size of the image window based on the radius of the region where the preset common feature of the wheel is located comprises:

selecting an image window having a width of w=2kr and a length of l=3kr; and

determining the position of the image window based on the radius of the region where the preset common feature of the wheel is located and the center of the region where the preset common feature of the wheel is located comprises;

determining the position of the image window, with the position of the image window being within a value range of [x−

6r, x+6r] in a horizontal direction and within a value range of [y−

3r, y+3r] in a vertical direction,

wherein k is a scaling factor, a value range of k is greater than or equal to 1 and less than or equal to 3, r is a radius of the region where the preset common feature of the wheel is located, and x and y are coordinates of the center of the region where the preset common feature of the wheel is located.