WO2020155998A1 - 识别车体方向的方法和装置 - Google Patents

识别车体方向的方法和装置 Download PDF

Info

Publication number
WO2020155998A1
WO2020155998A1 PCT/CN2019/129805 CN2019129805W WO2020155998A1 WO 2020155998 A1 WO2020155998 A1 WO 2020155998A1 CN 2019129805 W CN2019129805 W CN 2019129805W WO 2020155998 A1 WO2020155998 A1 WO 2020155998A1
Authority
WO
WIPO (PCT)
Prior art keywords
vehicle
image
vehicle body
picture
body direction
Prior art date
Application number
PCT/CN2019/129805
Other languages
English (en)
French (fr)
Inventor
蒋晨
Original Assignee
阿里巴巴集团控股有限公司
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 阿里巴巴集团控股有限公司 filed Critical 阿里巴巴集团控股有限公司
Publication of WO2020155998A1 publication Critical patent/WO2020155998A1/zh

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/24—Classification techniques
    • G06F18/241—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V2201/00—Indexing scheme relating to image or video recognition or understanding
    • G06V2201/08—Detecting or categorising vehicles

Definitions

  • One or more embodiments of this specification relate to the field of computers, and in particular to methods and devices for identifying the direction of a vehicle body.
  • the intelligent image damage assessment refers to the process of taking pictures of the loss situation of the vehicle in danger by the artificial intelligence algorithm taken by ordinary users, and automatically identifying the lost parts and the degree of damage and automatically generating the vehicle maintenance plan.
  • One or more embodiments of this specification describe a method and device for identifying the direction of a vehicle body, which can efficiently identify the direction of a vehicle body.
  • a method for identifying the direction of a vehicle body is provided.
  • the method is used in a vehicle damage assessment case.
  • the method includes:
  • the vehicle body direction recognition of a single image is performed for each of the multiple vehicle images to obtain the vehicle body direction confidence vector of the single image, and the vehicle body direction confidence vector of the single image is used to represent the vehicle image
  • the target is determined at least according to the vehicle body direction confidence vector of the single image of the target vehicle image and the detection result of the body part of the single image of the target vehicle image
  • the body direction of the vehicle picture including:
  • the vehicle body direction of the target vehicle picture is determined according to the vehicle body direction confidence vector of the single image of the target vehicle picture and the component matching corresponding relationship of each vehicle picture.
  • the vehicle body direction category includes any one of the following:
  • the performing single-image vehicle body part recognition for each of the multiple vehicle pictures to obtain a single-image vehicle body part detection result includes:
  • Each vehicle picture in the plurality of vehicle pictures is used as the input of the pre-trained second neural network model to obtain the vehicle body part detection result of the single image of each vehicle picture, wherein the second neural network model adopts Target detection algorithm.
  • performing component matching calculation on the multiple vehicle pictures pairwise to determine the component matching correspondence of each vehicle picture include:
  • the multiple vehicle pictures are paired by component matching calculation to determine the component matching corresponding relationship of each vehicle picture.
  • the determining the vehicle body direction of the target vehicle picture according to the vehicle body direction confidence vector of the single image of the target vehicle picture and the component matching corresponding relationship of each vehicle picture includes:
  • the vehicle body direction confidence vector of the single image of the target vehicle image and the component matching corresponding relationship of each vehicle image are used as the input of a pre-trained decision model to obtain the vehicle body direction of the target vehicle image.
  • decision model adopts any of the following algorithms:
  • a device for identifying the direction of a vehicle body is provided.
  • the device is used in a vehicle damage assessment case, and the device includes:
  • the acquisition unit is used to acquire multiple vehicle pictures of the same damage assessment case
  • the single-image vehicle body direction recognition unit is configured to perform single-image vehicle body direction recognition for each of the multiple vehicle pictures acquired by the acquisition unit to obtain the single-image vehicle body direction confidence vector, and the single image
  • the car body direction confidence vector of the graph is used to represent the possibility that the car body direction of the vehicle picture belongs to each car body direction category;
  • the single-image vehicle body part recognition unit is configured to perform single-image vehicle body part recognition for each of the multiple vehicle images acquired by the acquisition unit to obtain a single-image vehicle body part detection result;
  • the determining unit is configured to at least according to the vehicle body direction confidence vector of the single image of the target vehicle image obtained by the single image vehicle body direction recognition unit and the target vehicle picture obtained by the single image vehicle body component recognition unit The vehicle body part detection result of the single image determines the vehicle body direction of the target vehicle picture.
  • a computer-readable storage medium on which a computer program is stored, and when the computer program is executed in a computer, the computer is caused to execute the method of the first aspect.
  • a computing device including a memory and a processor, the memory stores executable code, and the processor implements the method of the first aspect when the executable code is executed by the processor.
  • the vehicle body direction recognition of a single image is performed for each of the multiple vehicle pictures to obtain a single image
  • Vehicle body direction confidence vector the vehicle body direction confidence vector of the single image is used to characterize the possibility that the vehicle body direction of the vehicle picture belongs to each vehicle body direction category
  • Car body part recognition of the single image is performed on the vehicle image, and the detection result of the single image is obtained.
  • the vehicle body direction confidence vector of the single image of the target vehicle image and the single image of the target vehicle image The vehicle body component detection result of the, determines the vehicle body direction of the target vehicle picture. It can be seen from the above that in the embodiments of this specification, the direction recognition of the vehicle picture is combined with the component recognition to determine the body direction of the final vehicle picture, which can efficiently recognize the vehicle body direction, and the result has a high degree of confidence.
  • Fig. 1 is a schematic diagram of an implementation scenario of an embodiment disclosed in this specification
  • Figure 2 shows a flow chart of a method for identifying the direction of a vehicle body according to an embodiment
  • FIG. 3 shows a schematic diagram of the detection result of a single image of a vehicle body part according to an embodiment
  • Fig. 4 shows a schematic block diagram of an apparatus for identifying the direction of a vehicle body according to an embodiment
  • Fig. 5 shows a schematic block diagram of an apparatus for identifying the direction of a vehicle body according to another embodiment.
  • FIG. 1 is a schematic diagram of an implementation scenario of an embodiment disclosed in this specification.
  • This implementation scenario involves recognizing the direction of the vehicle body in the vehicle picture.
  • automatic loss determination is performed based on multiple vehicle pictures taken by the user.
  • various shooting angles such as vertical, horizontal, and oblique shots often appear, resulting in inconsistencies in the direction of the vehicle body.
  • the vehicle body direction is to be recognized for multiple vehicle pictures used for vehicle damage assessment, and the vehicle body direction of each vehicle picture is recognized for subsequent automatic damage assessment.
  • Figure 1 shows a typical example of possible car body directions, where Figure 1 (a) represents the direction of the car body upward, Figure 1 (b) represents the direction of the car body downward, and Figure 1 (c) Represents the direction of the car body to the left, and Figure 1(d) represents the direction of the car body to the right. It is understandable that there are special marking specifications for up, down, left, and right. Generally, the direction is in line with people's common sense. For example, the direction of the earth is downward, and the direction of the sky is upward.
  • multiple vehicle pictures used for vehicle damage assessment belong to the same damage assessment case.
  • multiple vehicle pictures include a long-range picture, a close-range picture, and a middle-range picture.
  • vehicle body direction recognition is beneficial to improve the accuracy of vehicle body direction recognition.
  • Fig. 2 shows a flow chart of a method for identifying the direction of a vehicle body according to an embodiment.
  • the method is used in a vehicle damage assessment case.
  • the method for identifying the direction of the vehicle body in this embodiment includes the following steps: Step 21: Obtain multiple vehicle pictures of the same damage assessment case; Step 22: For each vehicle picture in the multiple vehicle pictures Carry out the vehicle body direction recognition of the single image to obtain the vehicle body direction confidence vector of the single image.
  • the vehicle body direction confidence vector of the single image is used to represent the possibility that the vehicle body direction of the vehicle picture belongs to each vehicle body direction category
  • Step 23 For each of the multiple vehicle pictures, perform single-image body part recognition to obtain single-image body part detection results; Step 24, at least according to the single-image of the target vehicle image
  • the vehicle body direction confidence vector and the vehicle body part detection result of the single image of the target vehicle picture determine the vehicle body direction of the target vehicle picture.
  • step 21 obtain multiple vehicle pictures of the same damage assessment case. It is understandable that in the same damage assessment case, the vehicle images contained in multiple vehicle pictures belong to the same vehicle.
  • the vehicle picture can include the entire vehicle, for example, a long-range photo; the vehicle picture can also include a part of the vehicle, for example, a close-up photo or a middle-range photo.
  • multiple vehicle pictures may include both long-range photos, close-range photos and mid-range photos.
  • the user can obtain a video containing vehicle images by shooting a video, and then obtain multiple vehicle pictures by extracting multiple video frames of the video; or, the user can directly take multiple photos containing vehicle images, and then this Multiple photos as multiple vehicle pictures.
  • step 22 the vehicle body direction recognition of the single image is performed for each of the multiple vehicle pictures to obtain the vehicle body direction confidence vector of the single image, and the vehicle body direction confidence vector of the single image is used as In order to characterize the possibility of the vehicle body direction of the vehicle picture belonging to each vehicle body direction category. It is understandable that this process is performed separately for each vehicle picture, and different vehicle pictures do not affect each other's recognition results regarding the direction of the vehicle body.
  • each vehicle picture in the plurality of vehicle pictures is used as the input of the pre-trained first neural network model to obtain the vehicle body direction confidence vector of the single image of each vehicle picture, wherein
  • the first neural network model is a classifier.
  • vehicle body direction category includes any one of the following: up, down, left, right, unjudgeable.
  • step 23 the single-image vehicle body part recognition is performed for each of the multiple vehicle images to obtain the single-image vehicle body part detection result. It is understandable that this process is performed separately for each vehicle picture, and different vehicle pictures do not affect each other's detection results on vehicle body parts.
  • the detection result of the vehicle body part of the single image includes the category and position of the part (ie, the part area).
  • the category of the vehicle component is the headlight 31, and the position of the vehicle component is the component area enclosed by the rectangular frame 32. It is understandable that one or more vehicle components can be identified in a vehicle picture.
  • each vehicle picture in the plurality of vehicle pictures is used as the input of the pre-trained second neural network model to obtain the vehicle body part detection result of the single image of each vehicle picture, wherein the first Second, the neural network model uses the target detection algorithm.
  • the body direction of the target vehicle picture is determined at least according to the vehicle body direction confidence vector of the single image of the target vehicle picture and the detection result of the body part of the single image of the target vehicle picture . It is understandable that, in combination with the detection results of vehicle body parts, according to the relative position relationship between the parts, it is helpful to improve the accuracy of vehicle body direction recognition.
  • the multiple vehicle pictures are paired by part matching calculation to determine the part matching correspondence of each vehicle picture ; And then according to the vehicle body direction confidence vector of the single image of the target vehicle image and the component matching corresponding relationship of each vehicle image to determine the vehicle body direction of the target vehicle image. That is to say, when recognizing the vehicle body direction of a vehicle picture, combining the component detection results of multiple vehicle pictures helps to further improve the accuracy of vehicle body direction recognition.
  • the part recognition based on a single image may be unreliable, that is, it may be judged from the single image, and it is difficult to determine which part it is, but it can be determined by combining the front and back images. Because some of the front and back pictures may be mid-range pictures, and some are close-up pictures, the close-up pictures sometimes don’t know what parts are based on a single picture, but the middle-range pictures are often easier to identify which parts are, so that you can go directly based on the feature information of the picture. Match the features and know the parts in the close-up image.
  • the component matching relationship of each vehicle picture can be determined in the following manner: first, according to the detection result of the body component of the single image of each vehicle picture in the multiple vehicle pictures, the feature description of each component in each vehicle picture is determined Vector; and then according to the feature description vector of each component in each vehicle picture, the multiple vehicle pictures are paired by component matching calculation to determine the component matching correspondence of each vehicle picture.
  • each vehicle picture it is also possible to determine the component matching relationship of each vehicle picture in the following way: first determine the feature description vector of each vehicle picture; then according to the feature description vector of each vehicle picture and the corresponding relationship of each vehicle picture in the multiple vehicle pictures According to the detection result of the vehicle body parts of the single image, the parts matching calculation is performed on the plurality of vehicle pictures pairwise to determine the part matching correspondence of each vehicle picture.
  • the features described by the feature description vector can be diverse.
  • the feature is the gray level of the picture.
  • the component matching is performed by calculating the difference of the feature area.
  • the component matching relationship of each vehicle picture can be specifically as shown in Table 1. Show.
  • Table 1 Matching relationship table of parts of multiple vehicle pictures
  • Part 11 in Picture 1 corresponds to Part 21 in Picture 2, and also corresponds to Part 31 in Picture 3. That is to say, part 11, part 21, and part 31 are essentially the same part, but appear in three different pictures of Picture 1, Picture 2 and Picture 3.
  • the part 12 in Picture 1 corresponds to Picture 3.
  • Component 32 that is, component 12 and component 32 are essentially the same component, but appear in two different pictures, picture 1 and picture 3; among them, component 13 in picture 1 corresponds to component 23 in picture 2. That is to say, the part 13 and the part 23 are essentially the same part, but appear in two different pictures, picture 1 and picture 2.
  • multiple algorithms can be used to match components between different pictures, for example, to calculate the features of the pictures, and then to calculate indicators such as the similarity of the features to find areas with similar features, so as to achieve component matching.
  • the vehicle body orientation confidence vector of the single image of the target vehicle image and the component matching relationship of each vehicle image are used as the input of a pre-trained decision model to obtain the vehicle body of the target vehicle image direction.
  • decision model adopts any one of the following algorithms: decision tree algorithm, support vector machine algorithm and random forest algorithm.
  • the vehicle body direction of the single image is used to represent the possibility that the vehicle body direction of the vehicle picture belongs to each vehicle body direction category, and then for each vehicle picture in the plurality of vehicle pictures Carry out the recognition of the car body parts of the single image to obtain the detection result of the car body parts of the single image.
  • the vehicle body direction confidence vector of the single image of the target vehicle image and the car body of the single image of the target vehicle image The body part detection result determines the body direction of the target vehicle picture.
  • a device for identifying the direction of a vehicle body and the device is used in a vehicle damage assessment case.
  • Fig. 4 shows a schematic block diagram of an apparatus for identifying the direction of a vehicle body according to an embodiment. As shown in FIG. 4, the device 400 includes:
  • the obtaining unit 41 is configured to obtain multiple vehicle pictures of the same damage assessment case
  • the single-image vehicle body direction recognition unit 42 is configured to perform single-image vehicle body direction recognition for each of the multiple vehicle pictures acquired by the acquisition unit 41 to obtain the single-image vehicle body direction confidence vector, so The vehicle body direction confidence vector of the single image is used to represent the possibility that the vehicle body direction of the vehicle picture belongs to each vehicle body direction category;
  • the single-image vehicle body part recognition unit 43 is configured to perform single-image vehicle body part recognition for each of the multiple vehicle pictures acquired by the acquisition unit 41 to obtain a single-image vehicle body part detection result;
  • the determining unit 44 is configured to at least according to the vehicle body direction confidence vector of the single image of the target vehicle image obtained by the single image vehicle body direction recognition unit 42 and the single image vehicle body component recognition unit 43
  • the vehicle body part detection result of the single image of the target vehicle picture determines the vehicle body direction of the target vehicle picture.
  • the determining unit 44 includes:
  • the matching subunit is used to perform component matching calculation on the multiple vehicle pictures pairwise according to the detection results of the body parts of the single image of each vehicle picture in the multiple vehicle pictures, and determine the component matching correspondence of each vehicle picture ;
  • the determining subunit is configured to determine the body of the target vehicle image according to the vehicle body direction confidence vector of the single image of the target vehicle image and the component matching corresponding relationship of each vehicle image determined by the matching subunit direction.
  • the vehicle body direction category includes any one of the following:
  • the single-image vehicle body direction recognition unit 42 is specifically configured to use each of the multiple vehicle images as the input of the pre-trained first neural network model to obtain The vehicle body direction confidence vector of the single image of each vehicle picture, wherein the first neural network model is a classifier.
  • the single-image vehicle body part recognition unit 43 is specifically configured to use each of the multiple vehicle pictures as the input of the pre-trained second neural network model to obtain The vehicle body part detection result of the single image of each vehicle picture, wherein the second neural network model adopts a target detection algorithm.
  • matching subunit is specifically used for:
  • the multiple vehicle pictures are paired by component matching calculation to determine the component matching corresponding relationship of each vehicle picture.
  • the determining subunit is specifically configured to use the vehicle body direction confidence vector of the single image of the target vehicle image and the component matching corresponding relationship of each vehicle image as the input of the pre-trained decision model to obtain State the body direction of the target vehicle picture.
  • decision model adopts any of the following algorithms:
  • the acquiring unit 41 first acquires multiple vehicle pictures of the same damage assessment case, and then the single-image vehicle body direction recognition unit 42 performs a single-image image for each of the multiple vehicle pictures.
  • Car body direction recognition to obtain the car body direction confidence vector of the single image, the car body direction confidence vector of the single image is used to represent the possibility that the car body direction of the vehicle picture belongs to each car body direction category, and then the single image
  • the vehicle body part recognition unit 43 performs single-image vehicle body part recognition for each of the multiple vehicle pictures to obtain the single-image vehicle body part detection result, and the final determination unit 44 at least according to the target vehicle picture
  • the vehicle body direction confidence vector of the single image and the vehicle body part detection result of the single image of the target vehicle picture determine the vehicle body direction of the target vehicle picture. It can be seen from the above that, in the embodiment of this specification, the direction recognition of the vehicle picture is combined with the component recognition to determine the body direction of the final vehicle picture, which can efficiently recognize the vehicle body direction, and the result has a high
  • Fig. 5 shows a schematic block diagram of a device for identifying the direction of a vehicle body according to another embodiment, the device being used in a vehicle damage assessment case.
  • the device 500 for identifying the direction of the vehicle body in this embodiment has input and output, where input: all fixed-loss pictures of the same case (picture stream); output: each fixed-loss picture in the picture stream Direction of body parts.
  • the device 500 includes:
  • Single-image car body direction recognition model 51 Recognize the car body direction of a single image for all pictures in the same case, and obtain the confidence vector of the car body direction of the single image, which represents the possibility that the car body direction in each picture belongs to each direction category Sex [car body direction category: up, down, left, right, unable to judge] (direction recognition for the whole picture);
  • Single-picture car body part recognition model 52 Carry out single-picture car body part recognition for all pictures in the same case, and obtain the single-picture car body part detection result (box for identifying the part area);
  • Image stream component matching module 53 take the car body direction confidence vector of the single image and the component detection result (component box and direction confidence score) of the single image as input, and perform component matching calculation on the images in the image stream, and predict The box of the matching area and the corresponding direction confidence (score);
  • Car body part direction decision model 54 For each picture, get the matching result (matching area box and corresponding direction confidence score) with all other fixed-loss pictures, as the input of the decision model, and output each final picture
  • the recognition result of the car body direction further, the recognition result of the car body part direction of each part in each picture can be obtained.
  • the car body part direction is used to identify the direction of the part, for example, the car body part direction is used to identify Whether it is the left headlight or the right headlight, or the direction of the body part is used to identify whether it is the left front fender or the right front fender.
  • the model structure of the single-image vehicle body part recognition model includes but is not limited to Faster-RCNN, RFCN, SSD, YOLO and other target detection algorithms;
  • the single-image vehicle body direction recognition model can be based on various basic networks of neural networks Training classifiers, including but not limited to mobilenet, squeezenet, inception, resnet, etc.; image stream component matching modules, including but not limited to grayscale (template matching) and feature-based (extracting features to generate feature descriptors, but not limited to based on Features of deep neural network) matching algorithm; car body component direction decision model, including but not limited to decision tree, support vector machine, random forest, etc.
  • component detection is combined with vehicle body direction recognition, and based on the deep learning method, component detection is used to identify candidate regions with more precise positioning positions, instead of relying only on the information of the whole picture for direction judgment, and can minimize other The influence of interference improves the accuracy of vehicle body component orientation recognition.
  • This solution not only uses the information of the entire image, but also integrates the information of the front and rear components in the image stream, and can identify the direction of each part of the car body; adding component detection can more accurately locate the body part and reduce the interference of light and irrelevant components. Robustness is good.
  • a computer-readable storage medium having a computer program stored thereon, and when the computer program is executed in a computer, the computer is caused to execute the method described in conjunction with FIG. 2.
  • a computing device including a memory and a processor, the memory is stored with executable code, and when the processor executes the executable code, the implementation described in conjunction with FIG. 2 method.

Landscapes

  • Engineering & Computer Science (AREA)
  • Data Mining & Analysis (AREA)
  • Theoretical Computer Science (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Bioinformatics & Cheminformatics (AREA)
  • Bioinformatics & Computational Biology (AREA)
  • Artificial Intelligence (AREA)
  • Evolutionary Biology (AREA)
  • Evolutionary Computation (AREA)
  • Physics & Mathematics (AREA)
  • General Engineering & Computer Science (AREA)
  • General Physics & Mathematics (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Image Analysis (AREA)

Abstract

一种识别车体方向的方法和装置，方法包括：获取同一定损案件的多个车辆图片（21）；针对所述多个车辆图片中的每个车辆图片进行单图的车体方向识别，得到单图的车体方向置信度向量，所述单图的车体方向置信度向量用于表征该车辆图片的车体方向归属各个车体方向类别的可能性（22）；针对所述多个车辆图片中的每个车辆图片进行单图的车体部件识别，得到单图的车体部件检测结果（23）；至少根据目标车辆图片的所述单图的车体方向置信度向量和所述目标车辆图片的所述单图的车体部件检测结果，确定所述目标车辆图片的车体方向（24），从而能够做到高效识别车体方向，并且结果的置信度高。

Description

识别车体方向的方法和装置 技术领域
本说明书一个或多个实施例涉及计算机领域，尤其涉及识别车体方向的方法和装置。
背景技术
在车险定损场景，通常存在人工车辆定损和智能图像定损两种定损方式。其中，在车险定损场景，需要保险专业定损员根据出险车辆的损失部件及其损失程度给出车辆的维修方案,这一过程是人工车辆定损的过程。而智能图像定损指的人工智能算法根据普通用户拍摄提交出险车辆的损失情况拍摄图片,由对其进行损失部件及其损伤程度进行自动识别并自动产生车辆维修方案的过程。
在智能图像定损这种定损方式中，用户在拍摄现场和定损照片时，往往会出现竖拍，横拍，斜拍等各种拍摄角度的照片，这会影响定损的准确度，增加定损的成本，为了提高定损的准确度，需要将车体方向识别出来，以便于后续的自动定损。
现有技术中通常采用人工识别车体方向，这种方法较为低效，且周期较长。
因此，希望能有改进的方案，能够高效的识别车体方向。
发明内容
本说明书一个或多个实施例描述了一种识别车体方向的方法和装置，能够高效的识别车体方向。
第一方面，提供了一种识别车体方向的方法，所述方法用于车辆定损案件，方法包括：
获取同一定损案件的多个车辆图片；
针对所述多个车辆图片中的每个车辆图片进行单图的车体方向识别，得到单图的车体方向置信度向量，所述单图的车体方向置信度向量用于表征该车辆图片的车体方向归属各个车体方向类别的可能性；
针对所述多个车辆图片中的每个车辆图片进行单图的车体部件识别，得到单图的车体部件检测结果；
至少根据目标车辆图片的所述单图的车体方向置信度向量和所述目标车辆图片的所述单图的车体部件检测结果，确定所述目标车辆图片的车体方向。
在一种可能的实施方式中，所述至少根据目标车辆图片的所述单图的车体方向置信度向量和所述目标车辆图片的所述单图的车体部件检测结果，确定所述目标车辆图片的车体方向，包括：
根据所述多个车辆图片中各车辆图片的单图的车体部件检测结果，将所述多个车辆图片两两进行部件匹配计算，确定各车辆图片的部件匹配对应关系；
根据所述目标车辆图片的所述单图的车体方向置信度向量和各车辆图片的部件匹配对应关系，确定所述目标车辆图片的车体方向。
在一种可能的实施方式中，所述车体方向类别包括以下任意一项：
上、下、左、右、无法判断。
在一种可能的实施方式中，所述针对所述多个车辆图片中的每个车辆图片进行单图的车体方向识别，得到单图的车体方向置信度向量，包括：
将所述多个车辆图片中的每个车辆图片分别作为预先训练的第一神经网络模型的输入，得到每个车辆图片的单图的车体方向置信度向量，其中所述第一神经网络模型为分类器。
在一种可能的实施方式中，所述针对所述多个车辆图片中的每个车辆图片进行单图的车体部件识别，得到单图的车体部件检测结果，包括：
将所述多个车辆图片中的每个车辆图片分别作为预先训练的第二神经网络模型的输入，得到每个车辆图片的单图的车体部件检测结果，其中所述第二神经网络模型采用目标检测算法。
进一步地，所述根据所述多个车辆图片中各车辆图片的单图的车体部件检测结果，将所述多个车辆图片两两进行部件匹配计算，确定各车辆图片的部件匹配对应关系，包括：
根据所述多个车辆图片中各车辆图片的单图的车体部件检测结果，确定每个车辆图片中各部件的特征描述向量；
根据每个车辆图片中各部件的特征描述向量，将所述多个车辆图片两两进行部件匹配计算，确定各车辆图片的部件匹配对应关系。
进一步地，所述根据所述目标车辆图片的所述单图的车体方向置信度向量和各车辆图片的部件匹配对应关系，确定所述目标车辆图片的车体方向，包括：
将所述目标车辆图片的所述单图的车体方向置信度向量和各车辆图片的部件匹配对应关系作为预先训练的决策模型的输入，得到所述目标车辆图片的车体方向。
进一步地，所述决策模型采用如下任意一种算法：
决策树算法、支持向量机算法和随机森林算法。
第二方面，提供了一种识别车体方向的装置，所述装置用于车辆定损案件，装置包括：
获取单元，用于获取同一定损案件的多个车辆图片；
单图车体方向识别单元，用于针对所述获取单元获取的多个车辆图片中的每个车辆图片进行单图的车体方向识别，得到单图的车体方向置信度向量，所述单图的车体方向置信度向量用于表征该车辆图片的车体方向归属各个车体方向类别的可能性；
单图车体部件识别单元，用于针对所述获取单元获取的多个车辆图片中的每个车辆图片进行单图的车体部件识别，得到单图的车体部件检测结果；
确定单元，用于至少根据所述单图车体方向识别单元得到的目标车辆图片的所述单图的车体方向置信度向量和所述单图车体部件识别单元得到的所述目标车辆图片的所述单图的车体部件检测结果，确定所述目标车辆图片的车体方向。
第三方面，提供了一种计算机可读存储介质，其上存储有计算机程序，当所述计算机程序在计算机中执行时，令计算机执行第一方面的方法。
第四方面，提供了一种计算设备，包括存储器和处理器，所述存储器中存储有可执行代码，所述处理器执行所述可执行代码时，实现第一方面的方法。
通过本说明书实施例提供的方法和装置，首先获取同一定损案件的多个车辆图片，然后针对所述多个车辆图片中的每个车辆图片进行单图的车体方向识别，得到单图的车体方向置信度向量，所述单图的车体方向置信度向量用于表征该车辆图片的车体方向归属各个车体方向类别的可能性，接着针对所述多个车辆图片中的每个车辆图片进行单图的车体部件识别，得到单图的车体部件检测结果，最后至少根据目标车辆图片的所述单图的车体方向置信度向量和所述目标车辆图片的所述单图的车体部件检测结果，确定所述目标车辆图片的车体方向。由上可见，本说明书实施例中将车辆图片的方向识别与部 件识别相结合，确定最终的车辆图片的车体方向，能够做到高效识别车体方向，并且结果的置信度高。
附图说明
为了更清楚地说明本发明实施例的技术方案，下面将对实施例描述中所需要使用的附图作简单地介绍，显而易见地，下面描述中的附图仅仅是本发明的一些实施例，对于本领域普通技术人员来讲，在不付出创造性劳动的前提下，还可以根据这些附图获得其它的附图。
图1为本说明书披露的一个实施例的实施场景示意图；
图2示出根据一个实施例的识别车体方向的方法流程图；
图3示出根据一个实施例的单图的车体部件检测结果示意图；
图4示出根据一个实施例的识别车体方向的装置的示意性框图；
图5示出根据另一个实施例的识别车体方向的装置的示意性框图。
具体实施方式
下面结合附图，对本说明书提供的方案进行描述。
图1为本说明书披露的一个实施例的实施场景示意图。该实施场景涉及对车辆图片中车体方向的识别。通常地，智能图像定损过程中，基于用户拍摄的多个车辆图片来进行自动定损。用户在拍摄车辆图片时，往往会出现竖拍、横拍、斜拍等各种拍摄角度，导致车体方向的不一致。本说明书实施例中，要对用于车辆定损的多个车辆图片进行车体方向的识别，识别出每个车辆图片的车体方向，以用于后续的自动定损。
参照图1，图1中示出了可能出现的车体方向的典型示例，其中，图1(a)代表车体方向向上，图1(b)代表车体方向向下，图1(c)代表车体方向向左，图1(d)代表车体方向向右。可以理解的是，上下左右是有专门的标注规范的，一般的就是符合人认知常识的方向，比如地的方向就是向下，天空的方向就是向上。
此外，用于车辆定损的多个车辆图片属于同一定损案件，例如，多个车辆图片包含远景图片、近景图片和中景图片。
本说明书实施例中，将车体方向识别与车体部件识别结合起来，有利于提高车体方 向识别的准确性。
图2示出根据一个实施例的识别车体方向的方法流程图，所述方法用于车辆定损案件。如图2所示，该实施例中识别车体方向的方法包括以下步骤：步骤21，获取同一定损案件的多个车辆图片；步骤22，针对所述多个车辆图片中的每个车辆图片进行单图的车体方向识别，得到单图的车体方向置信度向量，所述单图的车体方向置信度向量用于表征该车辆图片的车体方向归属各个车体方向类别的可能性；步骤23，针对所述多个车辆图片中的每个车辆图片进行单图的车体部件识别，得到单图的车体部件检测结果；步骤24，至少根据目标车辆图片的所述单图的车体方向置信度向量和所述目标车辆图片的所述单图的车体部件检测结果，确定所述目标车辆图片的车体方向。下面描述以上各个步骤的具体执行方式。
首先在步骤21，获取同一定损案件的多个车辆图片。可以理解的是，在同一定损案件中，多个车辆图片中包含的车辆图像属于同一车辆。车辆图片可以包含车辆的整体，例如，远景照片；车辆图片也可以包含车辆的局部，例如，近景照片或中景照片。作为示例，多个车辆图片中可以既包含远景照片，也包含近景照片和中景照片。
本说明实施例中，对于多个车辆图片的获取方式不做限定。例如，可以由用户通过拍摄视频的方式获取包含车辆图像的视频，后续通过抽取该视频的多个视频帧得到多个车辆图片；或者，由用户直接拍摄多个包含车辆图像的照片，后续将这多个照片作为多个车辆图片。
接着在步骤22，针对所述多个车辆图片中的每个车辆图片进行单图的车体方向识别，得到单图的车体方向置信度向量，所述单图的车体方向置信度向量用于表征该车辆图片的车体方向归属各个车体方向类别的可能性。可以理解的是，这一过程是针对每个车辆图片分别进行的，不同的车辆图片之间互不影响各自的关于车体方向的识别结果。
在一个示例中，将所述多个车辆图片中的每个车辆图片分别作为预先训练的第一神经网络模型的输入，得到每个车辆图片的单图的车体方向置信度向量，其中所述第一神经网络模型为分类器。
其中，所述车体方向类别包括以下任意一项：上、下、左、右、无法判断。
然后在步骤23，针对所述多个车辆图片中的每个车辆图片进行单图的车体部件识别，得到单图的车体部件检测结果。可以理解的是，这一过程是针对每个车辆图片分别进行的，不同的车辆图片之间互不影响各自的关于车体部件的检测结果。
其中，所述单图的车体部件检测结果包括部件的类别和位置(即部件区域)，参见图3所示的单图的车体部件检测结果示意图，在该车辆图片中，识别出的一个车辆部件的类别为大灯31，该车辆部件的位置即矩形框32所包围的部件区域。可以理解的是，一个车辆图片中，可以识别出一个或多个车辆部件。
在一个示例中，将所述多个车辆图片中的每个车辆图片分别作为预先训练的第二神经网络模型的输入，得到每个车辆图片的单图的车体部件检测结果，其中所述第二神经网络模型采用目标检测算法。
最后在步骤24，至少根据目标车辆图片的所述单图的车体方向置信度向量和所述目标车辆图片的所述单图的车体部件检测结果，确定所述目标车辆图片的车体方向。可以理解的是，结合车体部件检测结果，根据部件间的相对位置关系，有助于提高车体方向识别的准确性。
在一个示例中，先根据所述多个车辆图片中各车辆图片的单图的车体部件检测结果，将所述多个车辆图片两两进行部件匹配计算，确定各车辆图片的部件匹配对应关系；再根据所述目标车辆图片的所述单图的车体方向置信度向量和各车辆图片的部件匹配对应关系，确定所述目标车辆图片的车体方向。也就是说，在识别一个车辆图片的车体方向时，结合了多个车辆图片的部件检测结果，有助于进一步提高车体方向识别的准确性。
可以理解的是，有的时候基于单图的部件识别可能是不够置信的，就是说可能从单图来判断，难以确定这是哪个部件，但是如果结合前后图片，就可以判定了。因为前后图片可能有的是中景图，有的是近景图，近景图有时候根据单图是不知道什么部件的，但是中景图往往比较容易识别是什么部件，这样就可以直接根据图片的特征信息，去匹配特征，知道近景图中的部件。
进一步地，可以采用如下方式确定各车辆图片的部件匹配对应关系：先根据所述多个车辆图片中各车辆图片的单图的车体部件检测结果，确定每个车辆图片中各部件的特征描述向量；再根据每个车辆图片中各部件的特征描述向量，将所述多个车辆图片两两进行部件匹配计算，确定各车辆图片的部件匹配对应关系。
或者，还可以采用如下方式确定各车辆图片的部件匹配对应关系：先确定每个车辆图片的特征描述向量；然后根据每个车辆图片的特征描述向量和所述多个车辆图片中各车辆图片的单图的车体部件检测结果，将所述多个车辆图片两两进行部件匹配计算，确定各车辆图片的部件匹配对应关系。
其中，特征描述向量所描述的特征可以是多样的，例如，特征为图片的灰度，通过计算特征区域的差值，来进行部件匹配，各车辆图片的部件匹配对应关系具体可以如表一所示。
表一：多个车辆图片的部件匹配对应关系表
图片1 部件11 图片2 部件21 图片3 部件31
图片1 部件12 / / 图片3 部件32
图片1 部件13 图片2 部件23 / /
参见表一，图片1中识别出三个部件，分别为部件11、部件12和部件13，其中，图片1中的部件11对应于图片2中的部件21，还对应于图片3中的部件31，也就是说，部件11、部件21和部件31实质为同一部件，只是出现在图片1、图片2和图片3这三个不同的图片中；其中，图片1中的部件12对应于图片3中的部件32，也就是说，部件12和部件32实质为同一部件，只是出现在图片1和图片3这两个不同的图片中；其中，图片1中的部件13对应于图片2中的部件23，也就是说，部件13和部件23实质为同一部件，只是出现在图片1和图片2这两个不同的图片中。
本说明书实施例中，可以采用多种算法来进行不同图片间的部件匹配，例如，计算图片的特征，然后去计算特征的相似度之类的指标，找到特征相似的区域，从而实现部件匹配。
在一个示例中，将所述目标车辆图片的所述单图的车体方向置信度向量和各车辆图片的部件匹配对应关系作为预先训练的决策模型的输入，得到所述目标车辆图片的车体方向。
进一步地，所述决策模型采用如下任意一种算法：决策树算法、支持向量机算法和随机森林算法。
通过本说明书实施例提供的方法，首先获取同一定损案件的多个车辆图片，然后针对所述多个车辆图片中的每个车辆图片进行单图的车体方向识别，得到单图的车体方向置信度向量，所述单图的车体方向置信度向量用于表征该车辆图片的车体方向归属各个车体方向类别的可能性，接着针对所述多个车辆图片中的每个车辆图片进行单图的车体部件识别，得到单图的车体部件检测结果，最后至少根据目标车辆图片的所述单图的车体方向置信度向量和所述目标车辆图片的所述单图的车体部件检测结果，确定所述目标 车辆图片的车体方向。由上可见，本说明书实施例中将车辆图片的方向识别与部件识别相结合，确定最终的车辆图片的车体方向，能够做到高效识别车体方向，并且结果的置信度高。
根据另一方面的实施例，还提供一种识别车体方向的装置，所述装置用于车辆定损案件。图4示出根据一个实施例的识别车体方向的装置的示意性框图。如图4所示，该装置400包括：
获取单元41，用于获取同一定损案件的多个车辆图片；
单图车体方向识别单元42，用于针对所述获取单元41获取的多个车辆图片中的每个车辆图片进行单图的车体方向识别，得到单图的车体方向置信度向量，所述单图的车体方向置信度向量用于表征该车辆图片的车体方向归属各个车体方向类别的可能性；
单图车体部件识别单元43，用于针对所述获取单元41获取的多个车辆图片中的每个车辆图片进行单图的车体部件识别，得到单图的车体部件检测结果；
确定单元44，用于至少根据所述单图车体方向识别单元42得到的目标车辆图片的所述单图的车体方向置信度向量和所述单图车体部件识别单元43得到的所述目标车辆图片的所述单图的车体部件检测结果，确定所述目标车辆图片的车体方向。
可选地，作为一个实施例，所述确定单元44，包括：
匹配子单元，用于根据所述多个车辆图片中各车辆图片的单图的车体部件检测结果，将所述多个车辆图片两两进行部件匹配计算，确定各车辆图片的部件匹配对应关系；
确定子单元，用于根据所述目标车辆图片的所述单图的车体方向置信度向量和所述匹配子单元确定的各车辆图片的部件匹配对应关系，确定所述目标车辆图片的车体方向。
可选地，作为一个实施例，所述车体方向类别包括以下任意一项：
上、下、左、右、无法判断。
可选地，作为一个实施例，所述单图车体方向识别单元42，具体用于将所述多个车辆图片中的每个车辆图片分别作为预先训练的第一神经网络模型的输入，得到每个车辆图片的单图的车体方向置信度向量，其中所述第一神经网络模型为分类器。
可选地，作为一个实施例，所述单图车体部件识别单元43，具体用于将所述多个车辆图片中的每个车辆图片分别作为预先训练的第二神经网络模型的输入，得到每个车辆图片的单图的车体部件检测结果，其中所述第二神经网络模型采用目标检测算法。
进一步地，所述匹配子单元，具体用于：
根据所述多个车辆图片中各车辆图片的单图的车体部件检测结果，确定每个车辆图片中各部件的特征描述向量；
根据每个车辆图片中各部件的特征描述向量，将所述多个车辆图片两两进行部件匹配计算，确定各车辆图片的部件匹配对应关系。
进一步地，所述确定子单元，具体用于将所述目标车辆图片的所述单图的车体方向置信度向量和各车辆图片的部件匹配对应关系作为预先训练的决策模型的输入，得到所述目标车辆图片的车体方向。
进一步地，所述决策模型采用如下任意一种算法：
决策树算法、支持向量机算法和随机森林算法。
通过本说明书实施例提供的装置，首先获取单元41获取同一定损案件的多个车辆图片，然后单图车体方向识别单元42针对所述多个车辆图片中的每个车辆图片进行单图的车体方向识别，得到单图的车体方向置信度向量，所述单图的车体方向置信度向量用于表征该车辆图片的车体方向归属各个车体方向类别的可能性，接着单图车体部件识别单元43针对所述多个车辆图片中的每个车辆图片进行单图的车体部件识别，得到单图的车体部件检测结果，最后确定单元44至少根据目标车辆图片的所述单图的车体方向置信度向量和所述目标车辆图片的所述单图的车体部件检测结果，确定所述目标车辆图片的车体方向。由上可见，本说明书实施例中将车辆图片的方向识别与部件识别相结合，确定最终的车辆图片的车体方向，能够做到高效识别车体方向，并且结果的置信度高。
图5示出根据另一个实施例的识别车体方向的装置的示意性框图，所述装置用于车辆定损案件。如图5所示，该实施例中识别车体方向的装置500具有输入和输出，其中，输入：同一案件的所有定损图片(图片流)；输出：图片流中每一张定损图片的车体部件方向。
该装置500包括：
单图车体方向识别模型51：针对同一案件中的所有图片进行单图的车体方向识别，得到单图的车体方向置信度向量，表征各张图片中车体方向归属各个方向类别的可能性【车体方向类别：上，下，左，右，无法判断】(针对全图进行方向识别)；
单图车体部件识别模型52：针对同一案件中的所有图片进行单图的车体部件识别， 得到单图的车体部件检测结果(识别部件区域的box)；
图片流部件匹配模块53：将单图的车体方向置信度向量和单图的部件检测结果(部件box和方向置信度score)作为输入，将图片流中的图片两两进行部件匹配计算，预测匹配区域的box以及对应的方向置信度(score)；
车体部件方向决策模型54：针对每一张图片，得到其与其他所有定损图片的匹配结果(匹配区域box及对应的方向置信度score)，作为决策模型的输入，输出最终每一张图片车体方向的识别结果，进一步地，还可以得到每一张图片中的各部件的车体部件方向的识别结果，车体部件方向用于标识部件的方向，例如，车体部件方向用于标识是左大灯还是右大灯，或者，车体部件方向用于标识是左前翼子板还是右前翼子板。
本说明书实施例中，将全图的方向识别与部件识别结合，并融合整案的图片流部件匹配，得到最终的部件方向，使得结果置信度更高。在本方案中，单图车体部件识别模型的模型结构，有且不仅限于Faster-RCNN,RFCN,SSD,YOLO等目标检测算法；单图车体方向识别模型可基于神经网络的各种基础网络训练分类器，有且不仅限于mobilenet，squeezenet，inception，resnet等；图片流部件匹配模块，有且不仅限于基于灰度(模板匹配)和基于特征(提取特征生成特征描述子，有且不仅限于基于深度神经网络的特征)的匹配算法；车体部件方向决策模型，有且不仅限于决策树，支持向量机，随机森林等。
本说明书实施例中，将部件检测与车身方向识别结合，基于深度学习的方法，用部件检测更精确的定位部位识别候选区域，而不仅仅依赖于全图的信息进行方向判断，能尽量减少其他干扰的影响，提高车体部件方向识别的准确性。本方案不仅利用全图的信息，还融合图片流中前后部件信息，对于车体各个部位的方向均能进行识别；增加部件检测，能够更加准确的定位车身部位，减少光照以及无关部件等干扰，鲁棒性好。
根据另一方面的实施例，还提供一种计算机可读存储介质，其上存储有计算机程序，当所述计算机程序在计算机中执行时，令计算机执行结合图2所描述的方法。
根据再一方面的实施例，还提供一种计算设备，包括存储器和处理器，所述存储器中存储有可执行代码，所述处理器执行所述可执行代码时，实现结合图2所描述的方法。
本领域技术人员应该可以意识到，在上述一个或多个示例中，本发明所描述的功能可以用硬件、软件、固件或它们的任意组合来实现。当使用软件实现时，可以将这些功能存储在计算机可读介质中或者作为计算机可读介质上的一个或多个指令或代码进行传输。
以上所述的具体实施方式，对本发明的目的、技术方案和有益效果进行了进一步详细说明，所应理解的是，以上所述仅为本发明的具体实施方式而已，并不用于限定本发明的保护范围，凡在本发明的技术方案的基础之上，所做的任何修改、等同替换、改进等，均应包括在本发明的保护范围之内。

Claims (18)

  1. 一种识别车体方向的方法，所述方法用于车辆定损案件，所述方法包括：
    获取同一定损案件的多个车辆图片；
    针对所述多个车辆图片中的每个车辆图片进行单图的车体方向识别，得到单图的车体方向置信度向量，所述单图的车体方向置信度向量用于表征该车辆图片的车体方向归属各个车体方向类别的可能性；
    针对所述多个车辆图片中的每个车辆图片进行单图的车体部件识别，得到单图的车体部件检测结果；
    至少根据目标车辆图片的所述单图的车体方向置信度向量和所述目标车辆图片的所述单图的车体部件检测结果，确定所述目标车辆图片的车体方向。
  2. 如权利要求1所述的方法，其中，所述至少根据目标车辆图片的所述单图的车体方向置信度向量和所述目标车辆图片的所述单图的车体部件检测结果，确定所述目标车辆图片的车体方向，包括：
    根据所述多个车辆图片中各车辆图片的单图的车体部件检测结果，将所述多个车辆图片两两进行部件匹配计算，确定各车辆图片的部件匹配对应关系；
    根据所述目标车辆图片的所述单图的车体方向置信度向量和各车辆图片的部件匹配对应关系，确定所述目标车辆图片的车体方向。
  3. 如权利要求1所述的方法，其中，所述车体方向类别包括以下任意一项：
    上、下、左、右、无法判断。
  4. 如权利要求1所述的方法，其中，所述针对所述多个车辆图片中的每个车辆图片进行单图的车体方向识别，得到单图的车体方向置信度向量，包括：
    将所述多个车辆图片中的每个车辆图片分别作为预先训练的第一神经网络模型的输入，得到每个车辆图片的单图的车体方向置信度向量，其中所述第一神经网络模型为分类器。
  5. 如权利要求1所述的方法，其中，所述针对所述多个车辆图片中的每个车辆图片进行单图的车体部件识别，得到单图的车体部件检测结果，包括：
    将所述多个车辆图片中的每个车辆图片分别作为预先训练的第二神经网络模型的输入，得到每个车辆图片的单图的车体部件检测结果，其中所述第二神经网络模型采用目标检测算法。
  6. 如权利要求2所述的方法，其中，所述根据所述多个车辆图片中各车辆图片的单图的车体部件检测结果，将所述多个车辆图片两两进行部件匹配计算，确定各车辆图片 的部件匹配对应关系，包括：
    根据所述多个车辆图片中各车辆图片的单图的车体部件检测结果，确定每个车辆图片中各部件的特征描述向量；
    根据每个车辆图片中各部件的特征描述向量，将所述多个车辆图片两两进行部件匹配计算，确定各车辆图片的部件匹配对应关系。
  7. 如权利要求2所述的方法，其中，所述根据所述目标车辆图片的所述单图的车体方向置信度向量和各车辆图片的部件匹配对应关系，确定所述目标车辆图片的车体方向，包括：
    将所述目标车辆图片的所述单图的车体方向置信度向量和各车辆图片的部件匹配对应关系作为预先训练的决策模型的输入，得到所述目标车辆图片的车体方向。
  8. 如权利要求7所述的方法，其中，所述决策模型采用如下任意一种算法：
    决策树算法、支持向量机算法和随机森林算法。
  9. 一种识别车体方向的装置，所述装置用于车辆定损案件，所述装置包括：
    获取单元，用于获取同一定损案件的多个车辆图片；
    单图车体方向识别单元，用于针对所述获取单元获取的多个车辆图片中的每个车辆图片进行单图的车体方向识别，得到单图的车体方向置信度向量，所述单图的车体方向置信度向量用于表征该车辆图片的车体方向归属各个车体方向类别的可能性；
    单图车体部件识别单元，用于针对所述获取单元获取的多个车辆图片中的每个车辆图片进行单图的车体部件识别，得到单图的车体部件检测结果；
    确定单元，用于至少根据所述单图车体方向识别单元得到的目标车辆图片的所述单图的车体方向置信度向量和所述单图车体部件识别单元得到的所述目标车辆图片的所述单图的车体部件检测结果，确定所述目标车辆图片的车体方向。
  10. 如权利要求9所述的装置，其中，所述确定单元，包括：
    匹配子单元，用于根据所述多个车辆图片中各车辆图片的单图的车体部件检测结果，将所述多个车辆图片两两进行部件匹配计算，确定各车辆图片的部件匹配对应关系；
    确定子单元，用于根据所述目标车辆图片的所述单图的车体方向置信度向量和所述匹配子单元确定的各车辆图片的部件匹配对应关系，确定所述目标车辆图片的车体方向。
  11. 如权利要求9所述的装置，其中，所述车体方向类别包括以下任意一项：
    上、下、左、右、无法判断。
  12. 如权利要求9所述的装置，其中，所述单图车体方向识别单元，具体用于将所述多个车辆图片中的每个车辆图片分别作为预先训练的第一神经网络模型的输入，得到每 个车辆图片的单图的车体方向置信度向量，其中所述第一神经网络模型为分类器。
  13. 如权利要求9所述的装置，其中，所述单图车体部件识别单元，具体用于将所述多个车辆图片中的每个车辆图片分别作为预先训练的第二神经网络模型的输入，得到每个车辆图片的单图的车体部件检测结果，其中所述第二神经网络模型采用目标检测算法。
  14. 如权利要求10所述的装置，其中，所述匹配子单元，具体用于：
    根据所述多个车辆图片中各车辆图片的单图的车体部件检测结果，确定每个车辆图片中各部件的特征描述向量；
    根据每个车辆图片中各部件的特征描述向量，将所述多个车辆图片两两进行部件匹配计算，确定各车辆图片的部件匹配对应关系。
  15. 如权利要求10所述的装置，其中，所述确定子单元，具体用于将所述目标车辆图片的所述单图的车体方向置信度向量和各车辆图片的部件匹配对应关系作为预先训练的决策模型的输入，得到所述目标车辆图片的车体方向。
  16. 如权利要求15所述的装置，其中，所述决策模型采用如下任意一种算法：
    决策树算法、支持向量机算法和随机森林算法。
  17. 一种计算机可读存储介质，其上存储有计算机程序，当所述计算机程序在计算机中执行时，令计算机执行权利要求1-8中任一项的所述的方法。
  18. 一种计算设备，包括存储器和处理器，所述存储器中存储有可执行代码，所述处理器执行所述可执行代码时，实现权利要求1-8中任一项的所述的方法。
PCT/CN2019/129805 2019-02-01 2019-12-30 识别车体方向的方法和装置 WO2020155998A1 (zh)

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
CN201910105020.8 2019-02-01
CN201910105020.8A CN110569865B (zh) 2019-02-01 2019-02-01 识别车体方向的方法和装置

Publications (1)

Publication Number Publication Date
WO2020155998A1 true WO2020155998A1 (zh) 2020-08-06

Family

ID=68772816

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2019/129805 WO2020155998A1 (zh) 2019-02-01 2019-12-30 识别车体方向的方法和装置

Country Status (3)

Country Link
CN (1) CN110569865B (zh)
TW (1) TWI722558B (zh)
WO (1) WO2020155998A1 (zh)

Families Citing this family (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN110569865B (zh) * 2019-02-01 2020-07-17 阿里巴巴集团控股有限公司 识别车体方向的方法和装置

Citations (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN106384100A (zh) * 2016-09-28 2017-02-08 武汉大学 一种基于部件的精细车型识别方法
US20180247416A1 (en) * 2017-02-27 2018-08-30 Dolphin AI, Inc. Machine learning-based image recognition of weather damage
CN109145928A (zh) * 2017-06-16 2019-01-04 杭州海康威视数字技术股份有限公司 一种基于图像的车头朝向识别方法及装置
CN109229109A (zh) * 2017-07-04 2019-01-18 百度在线网络技术（北京）有限公司 判断车辆行驶方向的方法、装置、设备和计算机存储介质
CN110569865A (zh) * 2019-02-01 2019-12-13 阿里巴巴集团控股有限公司 识别车体方向的方法和装置

Family Cites Families (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
JP6169366B2 (ja) * 2013-02-08 2017-07-26 株式会社メガチップス 物体検出装置、プログラムおよび集積回路
CN104615998B (zh) * 2015-02-15 2017-10-24 武汉大学 一种基于多视角的车辆检索方法
CN108319907A (zh) * 2018-01-26 2018-07-24 腾讯科技（深圳）有限公司 一种车辆识别方法、装置和存储介质
CN108985343B (zh) * 2018-06-22 2020-12-25 深源恒际科技有限公司 基于深度神经网络的汽车损伤检测方法和系统
CN109242006A (zh) * 2018-08-23 2019-01-18 阿里巴巴集团控股有限公司 基于车型分类的识别车辆损伤的方法及装置
CN109215027B (zh) * 2018-10-11 2024-05-24 平安科技（深圳）有限公司 一种基于神经网络的车辆定损方法、服务器及介质

Patent Citations (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN106384100A (zh) * 2016-09-28 2017-02-08 武汉大学 一种基于部件的精细车型识别方法
US20180247416A1 (en) * 2017-02-27 2018-08-30 Dolphin AI, Inc. Machine learning-based image recognition of weather damage
CN109145928A (zh) * 2017-06-16 2019-01-04 杭州海康威视数字技术股份有限公司 一种基于图像的车头朝向识别方法及装置
CN109229109A (zh) * 2017-07-04 2019-01-18 百度在线网络技术（北京）有限公司 判断车辆行驶方向的方法、装置、设备和计算机存储介质
CN110569865A (zh) * 2019-02-01 2019-12-13 阿里巴巴集团控股有限公司 识别车体方向的方法和装置

Also Published As

Publication number Publication date
TWI722558B (zh) 2021-03-21
CN110569865B (zh) 2020-07-17
TW202030639A (zh) 2020-08-16
CN110569865A (zh) 2019-12-13

Similar Documents

Publication Publication Date Title
US10198623B2 (en) Three-dimensional facial recognition method and system
CN105335722B (zh) 一种基于深度图像信息的检测系统及方法
CN111862296B (zh) 三维重建方法及装置、系统、模型训练方法、存储介质
WO2019232866A1 (zh) 人眼模型训练方法、人眼识别方法、装置、设备及介质
WO2019232862A1 (zh) 嘴巴模型训练方法、嘴巴识别方法、装置、设备及介质
US11837017B2 (en) System and method for face recognition based on dynamic updating of facial features
US10445602B2 (en) Apparatus and method for recognizing traffic signs
CN112287866B (zh) 一种基于人体关键点的人体动作识别方法及装置
WO2019033574A1 (zh) 电子装置、动态视频人脸识别的方法、系统及存储介质
CN108805016B (zh) 一种头肩区域检测方法及装置
CN107463865B (zh) 人脸检测模型训练方法、人脸检测方法及装置
JP2016062610A (ja) 特徴モデル生成方法及び特徴モデル生成装置
CN111860352B (zh) 一种多镜头车辆轨迹全跟踪系统及方法
CN103971112B (zh) 图像特征提取方法及装置
CN112287868B (zh) 一种人体动作识别方法及装置
CN112200056B (zh) 人脸活体检测方法、装置、电子设备及存储介质
CN112287867B (zh) 一种多摄像头的人体动作识别方法及装置
CN112784712B (zh) 一种基于实时监控的失踪儿童预警实现方法、装置
CN113298158A (zh) 数据检测方法、装置、设备及存储介质
WO2021046773A1 (zh) 人脸防伪检测方法、装置、芯片、电子设备和计算机可读介质
WO2020155998A1 (zh) 识别车体方向的方法和装置
CN113902932A (zh) 特征提取方法、视觉定位方法及装置、介质和电子设备
WO2021047249A1 (zh) 数据预测方法、装置、设备及计算机可读存储介质
CN117152625A (zh) 一种基于CoordConv和YOLOv5的遥感小目标识别方法、系统、设备及介质
CN114067277A (zh) 行人图像识别方法、装置、电子设备及存储介质

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: 19913050

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: 19913050

Country of ref document: EP

Kind code of ref document: A1