WO2015027649A1 - 一种多尺度模型车辆检测方法 - Google Patents

一种多尺度模型车辆检测方法 Download PDF

Info

Publication number
WO2015027649A1
WO2015027649A1 PCT/CN2013/090408 CN2013090408W WO2015027649A1 WO 2015027649 A1 WO2015027649 A1 WO 2015027649A1 CN 2013090408 W CN2013090408 W CN 2013090408W WO 2015027649 A1 WO2015027649 A1 WO 2015027649A1
Authority
WO
WIPO (PCT)
Prior art keywords
vehicle
vehicle detection
image
blocks
scale model
Prior art date
Application number
PCT/CN2013/090408
Other languages
English (en)
French (fr)
Inventor
王飞跃
李叶
Original Assignee
东莞中国科学院云计算产业技术创新与育成中心
中国科学院自动化研究所
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by 东莞中国科学院云计算产业技术创新与育成中心, 中国科学院自动化研究所 filed Critical 东莞中国科学院云计算产业技术创新与育成中心
Publication of WO2015027649A1 publication Critical patent/WO2015027649A1/zh

Links

Classifications

    • GPHYSICS
    • G06COMPUTING; CALCULATING OR COUNTING
    • G06VIMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V10/00Arrangements for image or video recognition or understanding
    • G06V10/20Image preprocessing
    • G06V10/32Normalisation of the pattern dimensions

Definitions

  • the present invention relates to the field of vehicle detection technology, and more particularly to a multi-scale model vehicle detection method.
  • Video-based vehicle detection technology is an important part of the intelligent transportation system book, providing vehicle information for many applications, such as traffic video surveillance systems, driver assistance systems, smart cars, and more.
  • Vehicles of different scales may exist in traffic scenarios, which is a challenging problem in vehicle detection methods.
  • Many methods use scaling vehicle models or scaling input images to detect vehicles of different scales.
  • the distance between the vehicle and the camera vehicle-camera distance
  • the resolution of the vehicle the vehicle characteristics are different at different resolutions
  • the shape of the vehicle has also changed (some parts of the vehicle are gradually invisible as the vehicle moves away from the camera, such as the roof, etc.).
  • the invention establishes a vehicle detection method based on a multi-scale model, which can solve the vehicle detection problem under different vehicle-camera distances.
  • the technical problem solved by the present invention is to provide a multi-scale model vehicle detection method, which can solve the vehicle detection problem under different vehicle-camera distances.
  • the technical solution of the present invention to solve the above technical problem is:
  • the method includes multi-scale model modeling, multi-scale model learning and vehicle detection; the multi-scale model modeling is constructed by using two or more different mixed image templates; the multi-scale model learning is from actual traffic images. Obtaining an image of the vehicle as a training pattern, learning an edge block, a texture block, a color block, a flatness block, and an image likelihood probability of the mixed image template; the vehicle detecting is to perform template matching on the traffic image by using the mixed image template, Thereby the vehicle object is detected.
  • the step S1 described in the multi-scale model is to use not less than two different mixed image templates
  • the vehicle pair under the camera distance has different scales and different characteristics
  • the indicated vehicle object is closest to the camera, ⁇ contains one or more image blocks of edge block, texture block, color block and flatness block; as the 2 increases, the farther the vehicle object is represented from the camera and the vehicle
  • the object is gradually blurred into a flat area, and other types of image blocks gradually become flatness blocks.
  • step S2 multi-scale model learning includes the following steps:
  • Step S2-1 intercepting the vehicle image from the actual traffic image as the training image, the number of the training images is not less than one;
  • Step S2-2 learning from all the training images by using the message mapping method, 7 ⁇ " The image likelihood probability of all edge blocks, texture blocks, color blocks, flatness blocks, and ' ⁇ 1 , 7 ⁇ ...,.
  • the step S3 is performed by the vehicle, including: utilizing Detecting one or more vehicle candidates; The vehicle detection scores of the vehicle candidates are calculated; the vehicle detection scores of the vehicle candidates are compared with a vehicle detection threshold, and if the vehicle detection score is greater than or equal to the vehicle detection threshold, the corresponding vehicle candidate is the detected vehicle object.
  • the edge block is represented by a GabOT wavelet primitive in a specific direction; the texture block is represented by a gradient histogram in a local rectangular region of the training image; the color block is represented by a color histogram in a partial rectangular region of the training image;
  • the flatness block is represented by a superimposed response value of a Gabor filter in one or more directions within a local rectangular region of the training image.
  • the image likelihood probability of ⁇ - L u N ⁇ is:
  • the number of image blocks (all edge blocks, texture blocks, color blocks, flatness blocks in the image block), is the image/based probability, is a reference distribution, and is the corresponding to the jth image block.
  • the coefficient, / is the distance between the jth image block and the image area, is the normalization constant
  • the vehicle detection score is: the calculation step of the vehicle detection threshold is: first, template matching is performed on all the training images by using ' ⁇ 1 , 7 ⁇ '..., the vehicle is detected, and the corresponding vehicle detection score is calculated; The vehicle detection threshold is then estimated using the vehicle detection scores for all of the training images.
  • the beneficial effects of the invention are:
  • the present invention uses a plurality of mixed image modes with different scales and different features for changes in vehicle resolution and characteristics at different vehicle-camera distances in traffic images.
  • the board constructs a multi-scale model to improve the vehicle detection accuracy under different vehicle-camera distances.
  • each hybrid image template incorporates a variety of vehicle features, improving vehicle detection accuracy and adapting the invention to a variety of weather conditions.
  • FIG. 1 is a vehicle image diagram of different vehicle-camera distances in a complex traffic scene of the present invention
  • Figure 2 is a partial training image diagram of the present invention
  • Figure 3 is a multi-scale model of the present invention
  • FIG. 4 is a diagram of vehicle detection results in a complex traffic scene according to the present invention
  • FIG. 5 is a diagram of vehicle detection results under a larger vehicle-camera distance according to the present invention.
  • the vehicle detection method of the present invention is divided into three main steps: multi-scale model modeling, multi-scale model learning, and vehicle detection.
  • the three steps are described in detail below.
  • Step S1 Multi-scale model modeling. Use no less than two different mixed image templates
  • ⁇ 1 - J, N ⁇ ⁇ constitute a multi-scale model, which respectively represents vehicle objects under different vehicle-camera distances, H ⁇ have different scales and different characteristics.
  • the indicated vehicle object is closest to the camera and includes one or more image blocks of edge block, texture block, color block and flatness block type;
  • the further indicated vehicle object is farther from the camera and the vehicle object is gradually blurred into a flat area, and other types of image blocks gradually become flatness blocks.
  • ⁇ ' ⁇ indicates that the vehicle object is farthest from the camera, ' ⁇ contains only one or more edge blocks and flatness Piece.
  • ⁇ 3 contains one or more edge blocks, flatness blocks. 1 shows vehicle objects (a), (b), and (c) respectively indicated by ⁇ , ⁇ 2, and ⁇ 3 in the embodiment of the present invention.
  • the edge block is represented by a Gabor wavelet primitive in a specific direction.
  • a Gabor wavelet primitive in 16 directions is used to represent different edge blocks.
  • only a Gabor wavelet primitive of not less than one direction is selected. Yes, not limited to 16 directions.
  • the length and width of the Gabor wavelet primitive in ⁇ is 25 image pixels
  • the length and width of the Gabor wavelet primitive in T 2 are 17 image pixels
  • the length and width of the Gabor wavelet primitive in T 3 It is 13 image pixels.
  • the length and width of the Gabor wavelet primitives herein are not less than one image pixel, and are not limited to 25, 17, or 13 image pixels.
  • the texture block is represented by a gradient histogram in a local rectangular area of the training image.
  • the gradient histogram is obtained by counting the Gabor filter response values in 16 directions in the local rectangular area of the training image. As long as the Gabor filter response value of not less than one direction is calculated, it is not limited to 16 directions.
  • the local rectangular length and width are 48 image pixels, and the local rectangular length and width in T 2 are 24 image pixels. Of course, the local rectangular length and width are not less than one image pixel, and are not limited to 48 or 24 image pixels.
  • the color block is represented by a color histogram in a partial rectangular area of the training image.
  • the color histogram is obtained by counting pixel values of three color channels of the HSV color space in the partial rectangular area of the training image.
  • other color spaces of the image area can also be counted here, not limited to the HSV color space, and are not limited to three color channels, as long as it is not less than one.
  • the partial rectangle has a length and width of 30 image pixels, and the partial rectangle has a length and a width of ⁇ 2 16 image pixels.
  • the local rectangular length and width are not less than one image pixel, and are not limited to 30 or 16 image pixels.
  • the flatness block is represented by a superimposed value of Gabor filter response values in one or more directions in a partial rectangular area of the training image, and the value obtained by superimposing the Gabor filter response values in 16 directions represents the leveling in the embodiment of the present invention.
  • the degree block of course, it is only necessary to superimpose the Gabor filter response value of not less than one direction, and is not limited to 16 directions.
  • ⁇ partial rectangular length and width of said image pixels 40, T 2 in the aspect of local rectangular image pixels 20, in the local rectangular Î ⁇ 3 aspect image is 10 pixels.
  • the local rectangular length and width are not less than one image pixel, and are not limited to 40, 20, or 10 image pixels.
  • step S2-1 the vehicle image is intercepted from the actual traffic image as a training image, and the number of training images is not less than one.
  • the embodiment of the present invention uses 20 training images (which principles are used to make trade-offs, and what is the difference).
  • Figure 2 shows a portion of the training image.
  • Step S2-2 learning all edge blocks, texture blocks, color blocks, and flatness blocks in the ⁇ 7 ⁇ '''' 7 ⁇ from the training image by using an Information Projection Principle ' 7 ⁇ ..., 7 ⁇ image likelihood probability.
  • Figure 3 shows the ⁇ , T 2 and learned in the embodiment of the present invention.
  • the image likelihood probability of the ⁇ 7 I - L ⁇ . ⁇ ⁇ ⁇ > is:
  • is the number of image blocks in 7 (all edge blocks, texture blocks, color blocks, flatness blocks in the image block), image/based probability, is a reference distribution
  • 1 ⁇ 2 is the jth image block Corresponding coefficient
  • / is between the jth image block and the image area ⁇ 1 ⁇ 2 Distance
  • Z is the normalization constant.
  • Step S3 the vehicle detects, performs template matching on the test traffic image by using the ', ..., , 7 ⁇ , detects one or more vehicle candidates, and calculates vehicle detection scores of the vehicle candidates.
  • the vehicle detection scores of these vehicle candidates are compared with a vehicle detection threshold. If the vehicle detection score is greater than or equal to the vehicle detection threshold, the corresponding vehicle candidate is the detected vehicle object.
  • the calculation formula of the vehicle detection score is: .
  • the calculating step of the vehicle detection threshold is: First, template matching is performed on all the training images by using the ' 1 ⁇ ', the vehicle in the training image is detected, and then the corresponding vehicle detection score is calculated.
  • the vehicle detection threshold is then estimated using the vehicle detection scores for all of the training images.
  • Figure 4 illustrates vehicle detection results on a test traffic image in accordance with an embodiment of the present invention.
  • Figure 5 will be shown in Figure 4 (a)

Landscapes

  • Engineering & Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • General Physics & Mathematics (AREA)
  • Multimedia (AREA)
  • Theoretical Computer Science (AREA)
  • Image Analysis (AREA)
  • Image Processing (AREA)

Abstract

本发明涉及到车辆检测技术领域,特别涉及到一种多尺度模型车辆检测方法。本发明包括多尺度模型建模、多尺度模型学习和车辆检测三个步骤;所述多尺度模型建模是利用两个以上不同的混合图像模板构建;所述的多尺度模型学习是从实际交通图像中汲取车辆图像作为训练图形,学习所述混合图像模板的边缘块、纹理块、颜色块、平整度块和图像似然概率;所述车辆检测是利用所述混合图像模板对交通图像进行模板匹配,从而检测出车辆对象。本发明具有适应多种天气条件、一定程度的车辆变形等优点,特别是能准确检测与摄像机不同距离的车辆;可以应用于视频中车辆的检测。

Description

一种多尺度模型车辆检测方法 技术领域
本发明涉及到车辆检测技术说领域, 特别涉及到一种多尺度模型车辆检测方 法。
背景技术
基于视频的车辆检测技术是智能交通系统书重要的一部分, 为许多应用提供 车辆信息, 如交通视频监控系统、 驾驶辅助系统、 智能车等。 在交通场景中可 能存在不同尺度的车辆, 这是车辆检测方法的一个挑战性的问题。 很多方法利 用缩放车辆模型或缩放输入图像来检测不同尺度的车辆。 但是在一幅交通图像 中随着车辆与摄像机距离 (车辆-摄像机距离) 的不同, 不仅车辆的尺度发生变 化, 车辆分辨率也发生了变化 (不同分辨率下车辆特征不同), 而且更严重的是 车辆外形也发生了变化 (车辆某些部件随着车辆远离摄像机而逐渐不可见, 如 车顶等), 此时若通过缩放同一个车辆模型或缩放输入图像的方法检测车辆, 将 不能获取准确的检测结果。 因此, 针对不同的车辆-摄像机距离, 研究鲁棒的车 辆检测方法仍然是个挑战性的问题。 本发明建立了一种基于多尺度模型的车辆 检测方法, 可以解决不同车辆 -摄像机距离下的车辆检测问题。
发明内容
本发明解决的技术问题在于提供一种多尺度模型车辆检测方法, 可以解决 不同车辆 -摄像机距离下的车辆检测问题。 本发明解决上述技术问题的技术方案是:
包括多尺度模型建模、 多尺度模型学习和车辆检测三个步骤; 所述多尺度 模型建模是利用两个以上不同的混合图像模板构建; 所述的多尺度模型学习是 从实际交通图像中汲取车辆图像作为训练图形, 学习所述混合图像模板的边缘 块、 纹理块、 颜色块、 平整度块和图像似然概率; 所述车辆检测是利用所述混 合图像模板对交通图像进行模板匹配, 从而检测出车辆对象。
所述的步骤 S1 多尺度模型建模是利用不少于两个的不同的混合图像模板
;且成多尺度模型,
Figure imgf000004_0001
摄像机距离下的车辆对 ^具有不同尺度和不同特征;
表示的车辆对象离摄像机的距离最近, ^包含一个或多个边缘块、纹理块、 颜色块和平整度块等类型的图像块; 随 2的增大, 表示的车辆对象离摄像机越远且车辆对象逐渐被模糊成平整 区域, 中其他类型的图像块逐渐变为平整度块。
表示的车辆对象离摄像机最远, Τ'Λ· '仅包含一个或多个边缘块和平整度块。 所述的步骤 S2多尺度模型学习, 包括以下步骤:
步骤 S2-1 , 从实际交通图像中截取车辆图像作为训练图像, 训练图像的数 量不少于 1幅; 步骤 S2-2 ,利用消息映射法从所述所有训练图像中学习 ,7^ " Ί 中的 所有边缘块、 纹理块、 颜色块、 平整度块及 '^17^…, 的图像似然概率。
所述的步骤 S3车辆检测, 包括: 利用
Figure imgf000004_0002
检测出一个或多个车辆 候选者; 计算这些车辆候选者的车辆检测得分; 将这些车辆候选者的车辆检测得分与车辆检测阈值进行比较, 若车辆检测 得分大于等于车辆检测阈值, 则相应的车辆候选者为被检测的车辆对象。 所述边缘块由特定方向的 GabOT小波基元表示; 所述纹理块由训练图像的 局部矩形区域内的梯度直方图表示; 所述颜色块由训练图像的局部矩形区域内 的颜色直方图表示; 所述平整度块由训练图像的局部矩形区域内一个或多个方 向的 Gabor滤波器的叠加响应值表示。 所述的 ^ - L u N≥ 的图像似然概率是:
其中 Λ是!;:中图像块 (图像块包含 中的所有边缘块、 纹理块、 颜色块、 平整 度块) 的数量, 是图像 /基于 的概率, 是一个参考分布, ½是 中第 j个图像块对应的系数, /是 中第 j个图像块与图像区域 ½之间的距离, 是归一化常数
1υ;
所述车辆检测得分为: 所述车辆检测阈值的计算步骤是: 首先, 利用 '^17^ '… 对所有所述训练图像进行模板匹配, 检测出车辆, 并计算相应的车辆检测得分; 然后, 利用所有所述训练图像的车辆检测得分估计车辆检测阈值。 本发明的有益效果有:
( 1 )在多尺度模型建模中,针对在交通图像中不同车辆 -摄像机距离下车辆 分辨率及特征的变化, 本发明使用多个带有不同尺度和不同特征的混合图像模 板构建多尺度模型, 提高不同车辆-摄像机距离下的车辆检测正确率。
(2) 在多尺度模型建模中, 每个混合图像模板融合了多种车辆特征, 提高 了车辆检测正确率, 并使本发明适应多种天气条件。
(3 ) 在车辆检测中, 利用所述多尺度模型从测试交通图像中检测车辆, 不 仅实现车辆定位, 也能对车辆轮廓等信息详细描述。 附图说明 下面结合附图对本发明进一步说明: 图 1 为本发明复杂交通场景中不同车辆-摄像机距离下的车辆图像图; 图 2 为本发明部分训练图像图; 图 3 为本发明多尺度模型中多个混合图像模板图; 图 4 为本发明复杂交通场景下的车辆检测结果图; 图 5 为本发明较大车辆-摄像机距离下的车辆检测结果图。 具体实施方式 如图所示, 本发明的得车辆检测方法分为三个主要步骤: 多尺度模型建模, 多尺度模型学习和车辆检测。 以下详细介绍这三个步骤。 步骤 S1 : 多尺度模型建模。 利用不少于两个的不同的混合图像模板
{ 1 - J, N^≥ 组成多尺度模型, 分别表示在不同车辆- 摄像机距离下的车辆对象, H Ί 具有不同尺度和不同特征。
表示的车辆对象离摄像机的距离最近, 包含一个或多个边缘块、纹理块、 颜色块和平整度块类型的图像块;
随 i的增大, 表示的车辆对象离摄像机越远且车辆对象逐渐被模糊成平整 区域, 中其他类型的图像块逐渐变为平整度块。
^'ν表示的车辆对象离摄像机最远, 'ν仅包含一个或多个边缘块和平整度 块。
本发明实施例以 N = 3为例, ^包含一个或多个边缘块、 纹理块、 颜色块和 平整度块, T2包含一个或多个边缘块、 纹理块、 颜色块和平整度块, Τ3包含一个 或多个边缘块、 平整度块。 图 1展示了本发明实施例中 ^、 Τ2和 Τ3分别表示的车 辆对象 (a)、 (b) 和 (c
所述边缘块由特定方向的 Gabor小波基元表示, 本发明实施例使用 16个方 向的 Gabor小波基元表示不同的边缘块, 当然此处只要选择不少于 1个方向的 Gabor小波基元即可, 不限于 16个方向。 本发明实施例中, ^中的 Gabor小波 基元的长宽为 25个图像像素, T2中的 Gabor小波基元的长宽为 17个图像像素, T3中的 Gabor小波基元的长宽为 13个图像像素。 当然此处 Gabor小波基元的长 宽只要选择不小于 1个图像像素即可, 不限于 25、 17、 13个图像像素。
所述纹理块由训练图像的局部矩形区域内的梯度直方图表示, 本发明实施 例通过统计训练图像的局部矩形区域内的 16个方向的 Gabor滤波响应值得到所 述梯度直方图, 当然此处只要计算不少于 1个方向的 Gabor滤波响应值即可, 不限于 16个方向。 本发明实施例中, ^中所述局部矩形长宽为 48个图像像素, T2中所述局部矩形长宽为 24个图像像素。 当然此处局部矩形长宽只要不小于 1 个图像像素即可, 不限于 48、 24个图像像素。
所述颜色块由训练图像的局部矩形区域内的颜色直方图表示, 本发明实施 例通过统计训练图像的局部矩形区域内的 HSV颜色空间的三个颜色通道的像素 值得到所述颜色直方图, 当然此处也可以统计图像区域的其它颜色空间, 不限 于 HSV颜色空间, 并且也不限于三个颜色通道, 只要不少于 1个即可。 本发明 实施例中, ^中所述局部矩形长宽为 30个图像像素, Τ2中所述局部矩形长宽为 16个图像像素。 当然此处局部矩形长宽只要不小于 1个图像像素即可, 不限于 30、 16个图像像素。 所述平整度块由训练图像的局部矩形区域内的一个或多个方向的 Gabor滤 波响应值的叠加值表示, 本发明实施例通过叠加 16个方向的 Gabor滤波响应值 得到的值表示所述平整度块, 当然此处只要叠加不少于 1个方向的 Gabor滤波 响应值即可, 不限于 16个方向。 本发明实施例中, ^中所述局部矩形长宽为 40 个图像像素, T2中所述局部矩形长宽为 20个图像像素, Τ3中所述局部矩形长宽 为 10个图像像素。 当然此处局部矩形长宽只要不小于 1个图像像素即可, 不限 于 40、 20、 10个图像像素。 步骤 S2: 多尺度模型学习包括以下步骤:
步骤 S2-1 , 从实际交通图像中截取车辆图像作为训练图像, 训练图像的数量 不少于 1幅。 本发明实施例使用了 20幅训练图像 (遵循何种原则进行取舍, 有 何区别)。 图 2展示了部分的训练图像。
步骤 S2-2, 利用消息映射法(Information Projection Principle)从所述训练图 像中学习所述 ^^ 7^ ''' ' '7^中的所有边缘块、 纹理块、 颜色块、 平整度块及 ' 7 ^…,7^的图像似然概率。 图 3展示了本发明实施例中学习出的 ^、 T2
Τ3
所述 {7 I - L Ζ .^ Ν^ Ν > 的图像似然概率是:
Figure imgf000008_0001
其中 \是7 中图像块 (图像块包含 中的所有边缘块、 纹理块、 颜色块、 平整度块) 的数量, 是图像 /基于 的概率, 是一个参考分布, ½ 是 中第 j个图像块对应的系数, /是 中第 j个图像块与图像区域 ^½之间的 距离, Z 是归一化常数。
步骤 S3 ,车辆检测,利用所述' , …, ,7^对测试交通图像进行模板匹配, 检测出一个或多个车辆候选者, 并计算这些车辆候选者的车辆检测得分。 将这 些车辆候选者的车辆检测得分与车辆检测阈值进行比较, 若车辆检测得分大于 等于车辆检测阈值, 则相应的车辆候选者为被检测的车辆对象。 所述车辆检测得分的计算公式为:
Figure imgf000009_0001
所述车辆检测阈值的计算步骤是: 首先, 利用所述'1 ^ ' 对所有所述训练图像进行模板匹配, 检测出 训练图像中的车辆, 然后计算相应的车辆检测得分。
然后, 利用所有所述训练图像的车辆检测得分估计车辆检测阈值。
图 4展示了本发明实施例在测试交通图像上的车辆检测结果。图 5将图 4 (a)
- (c) 中的虚线框中的图像区域及其检测结果放大显示。
以上是对本发明具体实施方式的描述, 并非对本发明保护范围的限制; 凡 依前述描述可得之等效方案, 均应包含在本发明的保护范围之内。

Claims

权 利 要 求 书
1、 一种多尺度模型车辆检测方法, 其特征在于: 包括多尺度模型建模、 多 尺度模型学习和车辆检测三个步骤; 所述多尺度模型建模是利用两个以上不同 的混合图像模板构建; 所述的多尺度模型学习是从实际交通图像中汲取车辆图 像作为训练图形, 学习所述混合图像模板的边缘块、 纹理块、 颜色块、 平整度 块和图像似然概率; 所述车辆检测是利用所述混合图像模板对交通图像进行模 板匹配, 从而检测出车辆对象。
2、 根据权利要求 1所述的车辆检测方法, 其特征在于:
所述的步骤 S1 多尺度模型建模是利用不少于两个的不同的混合图像模板
^ ^ N N≥ 2½&成多尺度模型, Τι , ^…, 分别表示在不同车辆- 摄像机距离下的车辆对象,
Figure imgf000010_0001
…,7^具有不同尺度和不同特征;
表示的车辆对象离摄像机的距离最近, ^包含一个或多个边缘块、纹理块、 颜色块和平整度块等类型的图像块; 随 2的增大, 表示的车辆对象离摄像机越远且车辆对象逐渐被模糊成平整 区域, 中其他类型的图像块逐渐变为平整度块。
表示的车辆对象离摄像机最远, 7'Λ'仅包含一个或多个边缘块和平整度块。 所述的步骤 S2多尺度模型学习, 包括以下步骤:
步骤 S2-1 , 从实际交通图像中截取车辆图像作为训练图像, 训练图像的数 量不少于 1幅; 步骤 S2-2,利用消息映射法从所述所有训练图像中学习 ,7^ ""Ίν中的 所有边缘块、 纹理块、 颜色块、 平整度块及' ^ 的图像似然概率。 所述的步骤 S3车辆检测, 包括: 利用' '…7^对测试交通图像进行模板匹配, 检测出一个或多个车辆 候选者;
计算这些车辆候选者的车辆检测得分;
将这些车辆候选者的车辆检测得分与车辆检测阈值进行比较, 若车辆检测 得分大于等于车辆检测阈值, 则相应的车辆候选者为被检测的车辆对象。
3、 根据权利要求 1所述的车辆检测方法, 其特征在于: 所述的边缘块由特 定方向的 Gabor小波基元表示; 所述纹理块由训练图像的局部矩形区域内的梯 度直方图表示; 所述颜色块由训练图像的局部矩形区域内的颜色直方图表示; 所述平整度块由训练图像的局部矩形区域内一个或多个方向的 Gabor滤波器的 叠加响应值表示。
4、 根据权利要求 2所述的车辆检测方法, 其特征在于: 所述的边缘块由特 定方向的 Gabor小波基元表示; 所述纹理块由训练图像的局部矩形区域内的梯 度直方图表示; 所述颜色块由训练图像的局部矩形区域内的颜色直方图表示; 所述平整度块由训练图像的局部矩形区域内一个或多个方向的 Gabor滤波器的 叠加响应值表示。
5、 根据权利要求 1至 4任一项所述的车辆检测方法, 其特征在于: 所述的 d I - .L 的图像似然概率是: ;
Ή ^ ,
其中 中图像块 (图像块包含 中的所有边缘块、 纹理块、 颜色块、 平整 度块) 的数量, Ρ( Ί )是图像 /基于 的概率, 0是一个参考分布, '½是 中第 j个图像块对应的系数, /是 中第 j个图像块与图像区域 ^之间的距离, ^¾是归一化常数 t
6、 根据权利要求 1至 4任一项所述的车辆检测方法, 其特征在于: 所述车
1υ;
辆检测得分为: ¾
7、 根据权利要求 5所述的车辆检测方法, 其特征在于: 所述车辆检测得分 为:
Figure imgf000012_0001
8、 根据权利要求 1至 4任一项所述的车辆检测方法, 其特征在于: 所述车 辆检测阈值的计算步骤是: 首先, 利用 Λί对所有所述训练图像进行模板匹配, 检测出车辆, 并计算相应的车辆检测得分; 然后, 利用所有所述训练图像的车辆检测得分估计车辆检测阈值。
9、 根据权利要求 5所述的车辆检测方法, 其特征在于: 所述车辆检测阈值 的计算步骤是: 首先, 利用' ^17^…,'7^7对所有所述训练图像进行模板匹配, 检测出车辆, 并计算相应的车辆检测得分; 然后, 利用所有所述训练图像的车辆检测得分估计车辆检测阈值。
10、 根据权利要求 6所述的车辆检测方法, 其特征在于: 所述车辆检测阈 值的计算步骤是: 首先, 利用 ' 7^…,'7^对所有所述训练图像进行模板匹配, 检测出车辆, 并计算相应的车辆检测得分; 然后, 利用所有所述训练图像的车辆检测得分估计车辆检测阈值。
PCT/CN2013/090408 2013-08-27 2013-12-25 一种多尺度模型车辆检测方法 WO2015027649A1 (zh)

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
CN201310379385.2 2013-08-27
CN201310379385.2A CN103473566B (zh) 2013-08-27 2013-08-27 一种基于多尺度模型的车辆检测方法

Publications (1)

Publication Number Publication Date
WO2015027649A1 true WO2015027649A1 (zh) 2015-03-05

Family

ID=49798410

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2013/090408 WO2015027649A1 (zh) 2013-08-27 2013-12-25 一种多尺度模型车辆检测方法

Country Status (2)

Country Link
CN (1) CN103473566B (zh)
WO (1) WO2015027649A1 (zh)

Cited By (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN112784629A (zh) * 2019-11-06 2021-05-11 株式会社理光 图像处理方法、装置和计算机可读存储介质

Families Citing this family (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN103870832B (zh) * 2014-03-21 2017-11-03 浙江宇视科技有限公司 车辆整体特征提取方法及车辆型号识别方法
CN104881681B (zh) * 2015-05-22 2018-08-14 浙江大学 基于混合图模型的图像序列类别标注方法
CN110610123A (zh) * 2019-07-09 2019-12-24 北京邮电大学 一种多目标车辆检测方法、装置、电子设备及存储介质

Citations (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20100158321A1 (en) * 2008-12-18 2010-06-24 University Of Central Florida Research Foundation, Inc Homography-based passive vehicle speed measuring
CN101916383A (zh) * 2010-08-25 2010-12-15 浙江师范大学 基于多摄像机的车辆检测跟踪识别系统
CN102646199A (zh) * 2012-02-29 2012-08-22 湖北莲花山计算机视觉和信息科学研究院 复杂场景中的车型识别方法

Family Cites Families (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
EP1938247A2 (en) * 2005-09-14 2008-07-02 University of Massachusetts Multi-scale enveloping spectrogram signal processing for condition monitoring and the like
CN1897015A (zh) * 2006-05-18 2007-01-17 王海燕 基于机器视觉的车辆检测和跟踪方法及系统
CN100485710C (zh) * 2006-12-22 2009-05-06 四川川大智胜软件股份有限公司 利用数字图像处理技术识别车辆类型的方法
CN102142090B (zh) * 2011-03-15 2013-03-13 中国科学技术大学 车辆检测方法和系统
CN102289948B (zh) * 2011-09-02 2013-06-05 浙江大学 高速公路场景下一种多特征融合的多车辆视频跟踪方法
CN103218621B (zh) * 2013-04-21 2015-12-02 北京航空航天大学 一种室外视频监控中多尺度车辆的识别方法

Patent Citations (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20100158321A1 (en) * 2008-12-18 2010-06-24 University Of Central Florida Research Foundation, Inc Homography-based passive vehicle speed measuring
CN101916383A (zh) * 2010-08-25 2010-12-15 浙江师范大学 基于多摄像机的车辆检测跟踪识别系统
CN102646199A (zh) * 2012-02-29 2012-08-22 湖北莲花山计算机视觉和信息科学研究院 复杂场景中的车型识别方法

Cited By (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN112784629A (zh) * 2019-11-06 2021-05-11 株式会社理光 图像处理方法、装置和计算机可读存储介质

Also Published As

Publication number Publication date
CN103473566B (zh) 2016-09-14
CN103473566A (zh) 2013-12-25

Similar Documents

Publication Publication Date Title
JP4637618B2 (ja) 車線認識装置
US20180189577A1 (en) Systems and methods for lane-marker detection
CN107103277B (zh) 一种基于深度相机和3d卷积神经网络的步态识别方法
US9928426B1 (en) Vehicle detection, tracking and localization based on enhanced anti-perspective transformation
Beyeler et al. Vision-based robust road lane detection in urban environments
CN105869120A (zh) 一种图像拼接的实时性优化方法
CN105335955A (zh) 对象检测方法和对象检测装置
CN111160291B (zh) 基于深度信息与cnn的人眼检测方法
US10984263B2 (en) Detection and validation of objects from sequential images of a camera by using homographies
US20150302611A1 (en) Vehicle dimension estimation from vehicle images
JP4797846B2 (ja) 区画線検出装置及び車線検出装置
WO2011065399A1 (ja) 走路認識装置、車両、走路認識方法及び走路認識プログラム
Liu et al. Vehicle detection and ranging using two different focal length cameras
CN113205447A (zh) 用于车道线识别的道路图片标注方法和装置
CN110992424B (zh) 基于双目视觉的定位方法和系统
WO2015027649A1 (zh) 一种多尺度模型车辆检测方法
CN104036229A (zh) 基于回归的主动外观模型初始化方法
CN109917359A (zh) 基于车载单目视觉的鲁棒车辆距离估计方法
JP5548212B2 (ja) 横断歩道標示検出方法および横断歩道標示検出装置
EP2926317B1 (en) System and method for detecting pedestrians using a single normal camera
WO2014205787A1 (zh) 一种基于混合图像模板的车辆检测方法
Geiger et al. Object flow: A descriptor for classifying traffic motion
US11941892B2 (en) Method and device for providing data for creating a digital map
CN104732222B (zh) 一种基于深度摄像头的多特征人体识别方法
US11087150B2 (en) Detection and validation of objects from sequential images of a camera by using homographies

Legal Events

Date Code Title Description
121 Ep: the epo has been informed by wipo that ep was designated in this application

Ref document number: 13892626

Country of ref document: EP

Kind code of ref document: A1

NENP Non-entry into the national phase

Ref country code: DE

122 Ep: pct application non-entry in european phase

Ref document number: 13892626

Country of ref document: EP

Kind code of ref document: A1