WO2022213336A1 - 车辆驾驶环境异常监测方法、装置、电子设备和存储介质 - Google Patents

车辆驾驶环境异常监测方法、装置、电子设备和存储介质 Download PDF

Info

Publication number
WO2022213336A1
WO2022213336A1 PCT/CN2021/086050 CN2021086050W WO2022213336A1 WO 2022213336 A1 WO2022213336 A1 WO 2022213336A1 CN 2021086050 W CN2021086050 W CN 2021086050W WO 2022213336 A1 WO2022213336 A1 WO 2022213336A1
Authority
WO
WIPO (PCT)
Prior art keywords
module
feature
behavior
detection model
vehicle
Prior art date
Application number
PCT/CN2021/086050
Other languages
English (en)
French (fr)
Inventor
高毅鹏
刘力铭
黄凯明
Original Assignee
深圳市锐明技术股份有限公司
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 深圳市锐明技术股份有限公司 filed Critical 深圳市锐明技术股份有限公司
Priority to PCT/CN2021/086050 priority Critical patent/WO2022213336A1/zh
Priority to CN202180000757.9A priority patent/CN113287120A/zh
Publication of WO2022213336A1 publication Critical patent/WO2022213336A1/zh

Links

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V20/00—Scenes; Scene-specific elements
    • G06V20/50—Context or environment of the image
    • G06V20/59—Context or environment of the image inside of a vehicle, e.g. relating to seat occupancy, driver state or inner lighting conditions
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/25—Fusion techniques
    • G06F18/253—Fusion techniques of extracted features
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/045—Combinations of networks
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/047—Probabilistic or stochastic networks
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/08—Learning methods
    • G06N3/082—Learning methods modifying the architecture, e.g. adding, deleting or silencing nodes or connections
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V40/00—Recognition of biometric, human-related or animal-related patterns in image or video data
    • G06V40/10—Human or animal bodies, e.g. vehicle occupants or pedestrians; Body parts, e.g. hands
    • G06V40/16—Human faces, e.g. facial parts, sketches or expressions
    • G06V40/161—Detection; Localisation; Normalisation
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V40/00—Recognition of biometric, human-related or animal-related patterns in image or video data
    • G06V40/10—Human or animal bodies, e.g. vehicle occupants or pedestrians; Body parts, e.g. hands
    • G06V40/16—Human faces, e.g. facial parts, sketches or expressions
    • G06V40/168—Feature extraction; Face representation

Definitions

  • the present application relates to the technical field of safety monitoring, and in particular, to a method, device, electronic device and storage medium for abnormal monitoring of a vehicle driving environment.
  • the embodiments of the present application provide a method, device, electronic device, and storage medium for monitoring abnormal behavior of a vehicle driving environment, which can solve the problem of the lack of effective monitoring of abnormal behavior in the driving environment of the vehicle in the prior art, and the inability to effectively guarantee the driving process.
  • Driver and passenger safety issues are included in the vehicle driving environment.
  • an embodiment of the present application provides a method for monitoring abnormality of a vehicle driving environment, including:
  • the processing of the in-vehicle image by the behavior detection model includes: acquiring key point features of persons in the in-vehicle image and their associated embedded values, where the associated embedded values are used to identify the difference between the key point features. degree of association; determine the behavior state of the person according to the key point feature and its associated embedded value; determine whether there is abnormal behavior in the vehicle based on the behavior state;
  • the abnormal behavior is reported to a designated terminal.
  • the behavior detection model includes a feature extraction module, an attention module, and a feature fusion module, and the behavior detection model further includes a double-head decoupling structure;
  • the step of acquiring the key point features of people in the in-vehicle image and their associated embedded values includes:
  • the output result of the feature fusion module is subjected to task regression by using the double-head decoupling structure to obtain the key point feature of the person and its associated embedded value.
  • the feature extraction module includes a first feature extraction sub-module, a second feature extraction sub-module and a third feature extraction sub-module
  • the attention module includes a channel attention sub-module module and spatial attention sub-module
  • the step of outputting the human body features of the person through the feature extraction module includes:
  • the step of inputting the human body feature to the attention module, and performing adaptive weighting on the human body feature by the attention module includes:
  • the step of performing feature fusion on the output result of the attention module by the feature fusion module includes:
  • the output result of the channel attention sub-module and the output result of the spatial attention sub-module are feature-fused by the feature fusion module.
  • the behavior detection model further includes a spatiotemporal graph convolution module, and the step of determining the behavior state of the person according to the key point feature and its associated embedded value, include:
  • the target feature is obtained
  • the spatiotemporal graph convolution module Inputting the target feature to a spatiotemporal graph convolution module for processing, and obtaining the behavior status of the person output by the spatiotemporal graph convolution module; wherein, the spatiotemporal graph convolution module is used for spatially performing spatial analysis on the target feature.
  • the first graph convolution of the dimension is convolved with the second graph of the time dimension, and the behavior state of the person is determined according to the result of the first graph convolution and the result of the second graph convolution.
  • the vehicle driving environment abnormality monitoring method further includes:
  • model optimization is performed on the behavior detection model that has been trained to obtain a target behavior detection model
  • the abnormal behavior is reported to the designated terminal.
  • the step of performing model optimization on the trained behavior detection model to obtain a target behavior detection model according to a preset algorithm includes:
  • model distillation is performed on the teacher network model and the student network model to obtain a target behavior detection model.
  • an embodiment of the present application provides a vehicle driving environment abnormality monitoring device, including:
  • an image acquisition unit used to acquire in-vehicle images in real time
  • the abnormality detection unit is configured to input the in-vehicle image into the behavior detection model that has been trained for processing, and obtain an abnormality detection result output by the behavior detection model; wherein, the behavior detection model has a
  • the processing includes: acquiring the key point features of the people in the in-vehicle image and their associated embedding values, where the associated embedding values are used to identify the degree of association between the key point features; according to the key point features and their associated embedding values , determine the behavior state of the person; based on the behavior state, determine whether there is abnormal behavior in the vehicle;
  • An abnormality reporting unit configured to report the abnormal behavior to a designated terminal if it is determined according to the abnormality detection result that there is abnormal behavior in the vehicle.
  • embodiments of the present application provide an electronic device, including a memory, a processor, and computer-readable instructions stored in the memory and executable on the processor, where the processor executes the computer When the instruction is readable, the vehicle driving environment abnormality monitoring method as described in the first aspect above is implemented.
  • an embodiment of the present application provides a computer-readable storage medium, where the computer-readable storage medium stores computer-readable instructions, and when the computer-readable instructions are executed by a processor, the above-mentioned first aspect is implemented The above-mentioned vehicle driving environment anomaly monitoring method.
  • an embodiment of the present application provides a computer-readable instruction product, which, when the computer-readable instruction product runs on an electronic device, causes the electronic device to execute the method for monitoring an abnormality of a vehicle driving environment as described in the first aspect.
  • the in-vehicle image is acquired in real time, and the in-vehicle image is input into the behavior detection model that has been trained for processing, and the abnormal detection result output by the behavior detection model is obtained, wherein the behavior detection model
  • the processing of the in-vehicle image includes: acquiring key point features of people in the in-vehicle image and their associated embedded values, where the associated embedded values are used to identify the degree of association between the key point features; point features and their associated embedded values to determine the behavior state of the person; based on the behavior state, determine whether there is abnormal behavior in the vehicle, and use the behavior detection model to accurately and effectively detect abnormal behavior of the people in the vehicle.
  • the abnormal detection result it is determined that there is abnormal behavior in the vehicle, and the abnormal behavior is reported to the designated terminal to realize the supervision of the behavior of the driver and passengers in the vehicle during the driving process, so that the operation platform or relevant law enforcement units can timely Take measures to avoid accidents, so as to effectively ensure the safety of drivers and passengers.
  • Fig. 1 is the realization flow chart of the vehicle driving environment abnormality monitoring method provided by the embodiment of the present application
  • FIG. 2 is a specific implementation flow chart of acquiring key point features and their associated embedded values in the vehicle driving environment abnormality monitoring method provided by the embodiment of the present application;
  • FIG. 3 is a flowchart of a specific implementation of extracting human body features through a feature extraction module in the method for monitoring abnormality of a vehicle driving environment provided by an embodiment of the present application;
  • FIG. 4 is a specific implementation flowchart of adaptive weighting performed by an attention module in the vehicle driving environment abnormality monitoring method provided by the embodiment of the present application;
  • Fig. 5 is a specific implementation flow chart of determining a behavior state in the vehicle driving environment abnormal monitoring method provided by the embodiment of the present application;
  • FIG. 6 is a schematic diagram of an allocation strategy of domain subsets in a method for monitoring abnormality in a vehicle driving environment provided by an embodiment of the present application;
  • Fig. 7 is a specific implementation flowchart of model optimization in the vehicle driving environment abnormality monitoring method provided by the embodiment of the present application.
  • FIG. 8 is a structural block diagram of a vehicle driving environment abnormality monitoring device provided by an embodiment of the present application.
  • FIG. 9 is a schematic diagram of an electronic device provided by an embodiment of the present application.
  • the term “if” may be contextually interpreted as “when” or “once” or “in response to determining” or “in response to detecting “.
  • the phrases “if it is determined” or “if the [described condition or event] is detected” may be interpreted, depending on the context, to mean “once it is determined” or “in response to the determination” or “once the [described condition or event] is detected. ]” or “in response to detection of the [described condition or event]”.
  • references in this specification to "one embodiment” or “some embodiments” and the like mean that a particular feature, structure or characteristic described in connection with the embodiment is included in one or more embodiments of the present application.
  • appearances of the phrases “in one embodiment,” “in some embodiments,” “in other embodiments,” “in other embodiments,” etc. in various places in this specification are not necessarily All refer to the same embodiment, but mean “one or more but not all embodiments” unless specifically emphasized otherwise.
  • the terms “including”, “including”, “having” and their variants mean “including but not limited to” unless specifically emphasized otherwise.
  • the vehicle driving environment abnormality monitoring method provided by the embodiment of the present application can be applied to a vehicle-mounted intelligent terminal.
  • This application is not only applicable to business vehicles such as taxis and buses, but also to situations such as hitchhikers, hitchhiking by relatives and friends, and school buses.
  • the embodiments of the present application do not limit any specific types of terminal devices.
  • FIG. 1 shows an implementation process of a method for monitoring abnormality in a driving environment of a vehicle provided by an embodiment of the present application.
  • the execution end of the embodiment of the present application is an in-vehicle intelligent terminal, and the method process includes steps S101 to S103.
  • the specific implementation principles of each step are as follows:
  • S101 Acquire in-vehicle images in real time.
  • the vehicle-mounted intelligent terminal is provided with a camera, and the camera of the vehicle-mounted intelligent terminal is used to capture the in-vehicle image in real time.
  • the above in-vehicle image is an image to be processed.
  • the in-vehicle video is acquired in real time
  • the in-vehicle video is composed of a series of frame video images
  • several frames of in-vehicle images are extracted from the in-vehicle video.
  • a video image with a specified number of frames is extracted from the in-vehicle video captured by the camera of the vehicle-mounted intelligent terminal, and then a predetermined number of video images are selected from the extracted video images with the specified number of frames according to a preset image selection algorithm.
  • the frame video image is used as the in-vehicle image to be processed, and the number of frames can be customized by the user.
  • 18 frames of video images are extracted from the in-vehicle video captured by the camera of the in-vehicle intelligent terminal, and then according to a preset image extraction algorithm, 10 frames of in-vehicle images are selected from the extracted multi-frame video images as the vehicle to be processed. image inside.
  • the process of extracting in-vehicle images with a specified number of frames from the in-vehicle video captured by the camera of the in-vehicle intelligent terminal is also the process of preliminary screening of in-vehicle images. Include human characteristics, etc. to determine.
  • facial feature point detection is performed on multiple frames of video images captured by the camera of the vehicle-mounted intelligent terminal, and the video images that do not contain human faces are screened out to obtain video images containing human faces, and then extracted according to a preset Algorithm, extract the video image of the specified number of frames from the video image containing the face, and use the preset image selection algorithm to perform quality selection on the video image of the specified number of frames, and select the video image with relatively better quality as the to-be-processed video image. interior image of the car.
  • the quality selection in the embodiment of the present application is to use a quality algorithm to judge the quality of the image, then output a corresponding numerical value, and determine the image with the best quality as the in-vehicle image to be processed through the comparison result of the numerical values corresponding to each image.
  • quality selection is performed by a preset image selection algorithm, which can reduce the amount of calculation, avoid redundancy, and select the video image with the best quality as the in-vehicle image to be processed, which is beneficial to the accuracy of subsequent processing.
  • the effectiveness of face detection processing is improved.
  • the preset extraction algorithm may be a random extraction algorithm, which randomly extracts a set number of video images from the preliminarily screened video images.
  • video images may also be screened according to preset image standards, and video images of a specified number of frames that meet the preset image standards are extracted as the in-vehicle images to be processed, where the preset image standards include clear images. degree, face integrity, the number of key parts of the human body, or the angle of the face, etc.
  • S102 Input the in-vehicle image into the behavior detection model that has been trained for processing, and obtain an abnormality detection result output by the behavior detection model.
  • the above behavior detection model is a deep learning neural network model, which can be obtained by using sample images of various known behavior categories as training sample sets, for example, sample images in the Kinetics behavior data set can be used.
  • the processing of the in-vehicle image by the behavior detection model includes: obtaining the key point features of the people in the in-vehicle image and their associated embedded values, and the associated embedded values are used
  • the behavior state of the person is determined according to the key point feature and its associated embedded value, and based on the behavior state, it is determined whether there is abnormal behavior in the vehicle.
  • the input of the above-mentioned trained behavior detection model is the in-vehicle images of several frames.
  • the above behavior detection model needs to obtain the key point features of people in each frame of in-vehicle images, and at the same time, obtain the associated embedded value corresponding to the key point feature, and use the associated embedded value to determine the degree of association between the key point features.
  • the number of persons in the in-vehicle image is determined according to the key point features of the persons in the in-vehicle image and their associated embedded values.
  • the distance between the associated embedded values corresponding to the different key point features is relatively close, and when different key point features belong to different people, the different key point features The distance between the corresponding associated embedded values is farther.
  • the difference between the associated embedded values corresponding to each key point feature is calculated, and the similar key point features are determined according to the difference value, and the similar key point features are classified into the same feature group.
  • Quantity determines the number of personnel.
  • key point features whose difference between associated embedded values is less than or equal to a preset difference threshold are determined as similar key point features.
  • the above-mentioned behavior detection model includes a feature extraction module, an attention module and a feature fusion module, and FIG. 2 shows the above-mentioned key point features of persons in the vehicle image obtained and their associated embedded values.
  • A1 Input the in-vehicle image to the feature extraction module, and output the human body features of the person through the feature extraction module.
  • the above-mentioned feature extraction module is used to extract and output the human body characteristics of the person in the above-mentioned in-vehicle image.
  • Human features include human skeletal features.
  • the above feature extraction module includes a Moblienet v3 network, and the human skeleton features of the driver and the passenger are obtained by using the output of the Moblienet v3 network.
  • the feature extraction module includes a first feature extraction sub-module, a second feature extraction sub-module and a third feature extraction sub-module. As shown in FIG. 3 , the above-mentioned feature extraction module outputs the above The steps for the human characteristics of the person, including:
  • A11 Output the first human body feature of the first level through the above-mentioned first feature extraction sub-module.
  • the above-mentioned first feature extraction sub-module is used to extract the features of the first level.
  • A12 Output the second human body feature of the second level through the above-mentioned second feature extraction sub-module.
  • the above-mentioned second feature extraction sub-module is used to extract the features of the second level.
  • A13 Output the third human body feature of the third level through the third feature extraction sub-module, wherein the first level, the second level, and the third level have a progressive relationship.
  • the above-mentioned third feature extraction sub-module is used to extract features of the third level.
  • the first feature extraction sub-module, the second feature extraction sub-module and the third feature extraction sub-module respectively extract features of different levels from the above in-vehicle image. Enriching feature information by acquiring features at different levels is beneficial to improve the accuracy of behavior detection model detection.
  • the first level represents shallow semantics
  • the second level represents mid-level semantics
  • the third level represents deep semantics
  • the first level, the second level, and the third level have a decreasing relationship
  • the first level represents deep semantics
  • the second level represents middle level semantics
  • the third level represents shallow semantics
  • the first level represents the first resolution, and the first feature extraction sub-module is used to extract the human body features of the first resolution; the second level represents the first resolution, and the second feature extraction sub-module is used to extract the human body features of the first resolution.
  • the module is used to extract the human body features of the second resolution; the above-mentioned third level represents the third resolution, and the above-mentioned third feature extraction sub-module is used to extract the human body features of the third resolution.
  • the first resolution is lower than the second resolution, and the second resolution is lower than the third resolution.
  • A2 Input the human body features to the attention module, and perform adaptive weighting on the human body features through the attention module.
  • the above-mentioned attention module uses an attention mechanism to control the weighting coefficient, and adaptively weights the human body features extracted by the above-mentioned feature extraction module according to the weighting coefficient.
  • the output of the above attention module includes adaptively weighted human features.
  • the above-mentioned attention module includes a channel attention sub-module and a spatial attention sub-module.
  • the above-mentioned steps of inputting the above-mentioned human body features to the above-mentioned attention module, and performing adaptive weighting on the above-mentioned human body characteristics through the above-mentioned attention module include:
  • A21 Perform channel adaptive weighting on the first human body feature, the second human body feature, and the third human body feature through the channel attention sub-module.
  • A22 Perform spatial adaptive weighting on the first human body feature, the second human body feature, and the third human body feature through the spatial attention submodule.
  • the channel attention sub-module and the spatial attention sub-module are used to squeeze and stimulate the first human body feature, the second human body feature and the third human body feature respectively, so as to realize the Adaptive weighting of human features and third human feature channels and spaces.
  • the first human body feature, the second human body feature, and the third human body feature are input to the attention module after feature splicing.
  • A3 Input the output result of the attention module to the feature fusion module, and perform feature fusion on the output result of the attention module through the feature fusion module.
  • the above feature fusion module is used to fuse the adaptively weighted human body features to facilitate further analysis and processing.
  • feature fusion is performed on the output result of the channel attention sub-module and the output result of the spatial attention sub-module through the feature fusion module.
  • A4 Use the double-head decoupling structure to perform task regression on the output result of the feature fusion module to obtain the key point features of the person and their associated embedded values.
  • the above behavior detection model further includes a double-head decoupling structure.
  • Coupling refers to the phenomenon in which two or more systems or two forms of motion influence each other through interaction and even combine.
  • Decoupling deals with problems by mathematically separating the two motions. There are dependencies between modules and there must be coupling.
  • the output result of the feature fusion module includes a heatmap of key points and associated embedded values
  • the task regression is performed by using the double-head decoupling structure.
  • Task regression refers to determining the label of the person to which the keypoint feature belongs.
  • the above key point heatmap uses the loss function focal loss for regression
  • the associated embedded value uses the loss function push loss combined with the loss function pull loss to achieve regression.
  • the double-head decoupling structure is used to reduce the coupling degree between the modules, so as to avoid serious information return coupling.
  • the three sub-feature extraction modules C2, C3, and C4 of the Moblienet v3 network are used to extract the shallow, middle, and deep human features, respectively, and then the channel attention sub-module and the spatial attention sub-module are used to extract C2, C3,
  • the human body features extracted by C4 are adaptively weighted by channel and space.
  • the feature fusion module is used to integrate the adaptively weighted human body features, and then the dual-head decoupling structure is used to perform task regression on the output results of the feature fusion module to obtain the key point features and their associated embedded values of the people in the above-mentioned in-vehicle images.
  • the behavior detection model further includes a spatiotemporal graph convolution module, and FIG. 5 shows the specific process of determining the behavior state of the person according to the key point feature and its associated embedded value.
  • the target feature is obtained by splicing the feature vector of the key point feature and the associated embedded value.
  • the spatiotemporal graph convolution module Input the target feature into the spatiotemporal graph convolution module for processing, and obtain the behavior state of the person output by the spatiotemporal graph convolution module.
  • the spatiotemporal graph convolution module is used to perform the first graph convolution of the spatial dimension and the second graph convolution of the temporal dimension on the target feature, according to the result of the first graph convolution and the second graph convolution of the The result of the graph convolution determines the behavioral state of the person.
  • the above-mentioned spatiotemporal graph convolution module includes three layers of spatiotemporal graph convolution.
  • the spatiotemporal graph convolution can be similar to the ordinary convolution.
  • the first graph convolution of the spatial dimension is performed first, that is, different subsets are multiplied by different key point feature vectors, and the same subset is used. Multiply by the same key point feature vector, and then perform the second graph convolution in the time dimension.
  • the softmax function is used to regress the behavior state of the occupants in the car. In this way, the prior structure of the key points of the human skeleton is effectively used, and the information transmission in the time dimension is also effectively used, so that the behavior status of the people in the car can be accurately judged, and the abnormal detection of the driving environment in the car can be improved. accuracy.
  • the above-mentioned personnel include drivers and passengers.
  • the above behavior states are divided into normal behavior states and abnormal behavior states.
  • the abnormality detection result output by the above behavior detection model includes the behavior state of the person. If the behavior state is an abnormal behavior state, the abnormality detection result also includes abnormal behavior. Unusual behavior includes passenger harassment and beating of driver, and driver harassment and beating of passenger.
  • a notification event is immediately generated according to the abnormal behavior and uploaded to the designated terminal.
  • the above-mentioned vehicle-mounted intelligent terminal is communicatively connected with a designated terminal.
  • the above-mentioned designated terminals include, but are not limited to, the intelligent terminals of the operation platform, the intelligent terminals of the regulatory authorities, and the mobile terminals of designated users.
  • the above-mentioned method for monitoring abnormality in the driving environment of a vehicle further includes: if it is determined that there is abnormal behavior in the vehicle according to the abnormality detection result, sending an audio alarm, and prompting the driver and passengers in the vehicle to stop through the audio alarm The above abnormal behavior.
  • the method for monitoring abnormality of the vehicle driving environment further includes:
  • C1 According to a preset algorithm, perform model optimization on the behavior detection model that has been trained to obtain a target behavior detection model.
  • model optimization is performed on the behavior detection model that has been trained.
  • the above-mentioned preset algorithm includes model distillation.
  • the above-mentioned step C1 includes:
  • C11 Obtain the target training sample set.
  • the number of samples and the type of samples in the target training sample set can be determined according to actual needs, which is not specifically limited in this application.
  • the training sample set of the above behavior detection model can also be directly used.
  • C12 Obtain a teacher network model and a student network model, wherein the teacher network model is the trained behavior detection model, and the student network model is the trained network model after pruning according to preset parameters Behavior detection model.
  • the above preset parameter may be the model sensitivity.
  • the above pruning refers to pruning the number of channels in the convolutional layer in the model.
  • the pruning rate is set to 0.4, and 40% of the channels of a single convolutional layer in the above trained behavior detection model are pruned.
  • C13 Based on the adversarial generation network and the training sample set, perform model distillation on the teacher network model and the student network model to obtain a target behavior detection model.
  • the idea of adversarial generation network is added in the process of model distillation, which can effectively ensure the accuracy of the model.
  • C2 Use the target behavior detection model to process the in-vehicle image to obtain an abnormality detection result output by the target behavior detection model.
  • the trained behavior detection model is determined as the original model and used as the teacher network, and the trained behavior detection model after automatic pruning based on sensitivity analysis is used as the student network for model distillation.
  • the idea of GAN is added to the distillation process.
  • model distillation we use the teacher network as the discriminant model and the student network as the generative model, and use the discriminant network to effectively supervise the generative network to iteratively optimize the model.
  • the algorithm has high accuracy and low computational cost. network model.
  • structured pruning and model distillation are used to further compress the model, so as to obtain an optimized target behavior detection model, and the in-vehicle image is processed by using the target behavior detection model to obtain the target behavior detection model.
  • the abnormal detection results output by the model can greatly reduce the calculation amount of the model while avoiding the loss of model accuracy, which can enhance the real-time performance of abnormal behavior monitoring.
  • the in-vehicle image is acquired in real time, and the in-vehicle image is input into the behavior detection model that has been trained for processing, and the abnormal detection result output by the behavior detection model is obtained, wherein the
  • the processing of the in-vehicle image by the behavior detection model includes: acquiring key point features of persons in the in-vehicle image and their associated embedded values, where the associated embedded values are used to identify the degree of association between the key point features; According to the key point feature and its associated embedded value, determine the behavior state of the person; based on the behavior state, determine whether there is abnormal behavior in the vehicle, and use the behavior detection model to accurately and effectively measure the behavior of the person in the vehicle.
  • Anomaly detection if it is determined that there is an abnormal behavior in the vehicle according to the abnormal detection result, the abnormal behavior will be reported to the designated terminal to realize the supervision of the behavior of the driver and passengers in the vehicle during the driving process, so as to facilitate the operation of the platform or related Law enforcement units can take timely measures to avoid accidents, thereby effectively ensuring the safety of drivers and passengers.
  • FIG. 8 shows a structural block diagram of the vehicle driving environment abnormality monitoring device provided by the embodiment of the present application. relevant part.
  • the vehicle driving environment abnormality monitoring device includes: an image acquisition unit 81, an abnormality detection unit 82, and an abnormality reporting unit 83, wherein:
  • an image acquisition unit 81 configured to acquire in-vehicle images in real time
  • the abnormality detection unit 82 is configured to input the in-vehicle image into the behavior detection model that has been trained for processing, and obtain an abnormality detection result output by the behavior detection model;
  • the processing includes: acquiring key point features of people in the in-vehicle image and their associated embedding values, where the associated embedding values are used to identify the degree of association between key point features; according to the key point features and their associated embedding values value, determine the behavior state of the person; based on the behavior state, determine whether there is abnormal behavior in the vehicle;
  • the abnormality reporting unit 83 is configured to report the abnormal behavior to a designated terminal if it is determined according to the abnormality detection result that there is abnormal behavior in the vehicle.
  • the above-mentioned behavior detection model includes a feature extraction module, an attention module and a feature fusion module, and the above-mentioned behavior detection model also includes a double-head decoupling structure; the above-mentioned abnormality detection unit 82 is specifically used for:
  • the output result of the feature fusion module is subjected to task regression by using the double-head decoupling structure to obtain the key point feature of the person and its associated embedded value.
  • the above-mentioned feature extraction module includes a first feature extraction sub-module, a second feature extraction sub-module and a third feature extraction sub-module
  • the above-mentioned attention module includes a channel attention sub-module and a spatial attention sub-module submodule
  • the above-mentioned steps of outputting the human body features of the person through the feature extraction module include:
  • the above-mentioned steps of inputting the human body features to the attention module, and performing adaptive weighting on the human body features by the attention module include:
  • the above-mentioned steps of performing feature fusion on the output result of the attention module through the feature fusion module include:
  • the output result of the channel attention sub-module and the output result of the spatial attention sub-module are feature-fused by the feature fusion module.
  • the above behavior detection model further includes a spatiotemporal graph convolution module, and the above anomaly detection unit 82 is also used for:
  • the target feature is obtained
  • the spatiotemporal graph convolution module Inputting the target feature to a spatiotemporal graph convolution module for processing, and obtaining the behavior status of the person output by the spatiotemporal graph convolution module; wherein, the spatiotemporal graph convolution module is used for spatially performing spatial analysis on the target feature.
  • the first graph convolution of the dimension is convolved with the second graph of the time dimension, and the behavior state of the person is determined according to the result of the first graph convolution and the result of the second graph convolution.
  • the vehicle driving environment abnormality monitoring device further includes:
  • a model optimization unit configured to perform model optimization on the trained behavior detection model according to a preset algorithm to obtain a target behavior detection model
  • the above-mentioned abnormality detection unit 82 is further configured to process the in-vehicle image by using the target behavior detection model to obtain the abnormality detection result output by the target behavior detection model;
  • the abnormality reporting unit 83 is further configured to report the abnormal behavior to the designated terminal if it is determined that there is abnormal behavior in the vehicle according to the abnormality detection result output by the target behavior detection model.
  • the above-mentioned model optimization unit includes:
  • the sample acquisition module is used to acquire the target training sample set
  • the optimization training module is used to obtain a teacher network model and a student network model, wherein the teacher network model is the behavior detection model that has been trained, and the student network model is the model after pruning according to preset parameters.
  • the trained behavior detection model is used to obtain a teacher network model and a student network model, wherein the teacher network model is the behavior detection model that has been trained, and the student network model is the model after pruning according to preset parameters.
  • the target model generation module is configured to perform model distillation on the teacher network model and the student network model based on the confrontation generation network and the training sample set to obtain a target behavior detection model.
  • the in-vehicle image is acquired in real time, and the in-vehicle image is input into the behavior detection model that has been trained for processing, and the abnormal detection result output by the behavior detection model is obtained, wherein the
  • the processing of the in-vehicle image by the behavior detection model includes: acquiring key point features of persons in the in-vehicle image and their associated embedded values, where the associated embedded values are used to identify the degree of association between the key point features; According to the key point feature and its associated embedded value, determine the behavior state of the person; based on the behavior state, determine whether there is abnormal behavior in the vehicle, and use the behavior detection model to accurately and effectively measure the behavior of the person in the vehicle.
  • Anomaly detection if it is determined that there is an abnormal behavior in the vehicle according to the abnormal detection result, the abnormal behavior will be reported to the designated terminal to realize the supervision of the behavior of the driver and passengers in the vehicle during the driving process, so as to facilitate the operation of the platform or related Law enforcement units can take timely measures to avoid accidents, thereby effectively ensuring the safety of drivers and passengers.
  • Embodiments of the present application further provide a computer-readable storage medium, where the computer-readable storage medium stores computer-readable instructions, and when the computer-readable instructions are executed by a processor, any one of the instructions shown in FIG. 1 to FIG. 7 is implemented.
  • the invention discloses the steps of a vehicle driving environment abnormality monitoring method.
  • Embodiments of the present application further provide an electronic device, including a memory, a processor, and computer-readable instructions stored in the memory and executable on the processor, when the processor executes the computer-readable instructions.
  • Embodiments of the present application further provide a computer-readable instruction product, which, when the computer-readable instruction product runs on an electronic device, enables the electronic device to execute any one of the methods for monitoring abnormality in the driving environment of a vehicle as shown in FIGS. 1 to 7 .
  • FIG. 9 is a schematic diagram of an electronic device provided by an embodiment of the present application.
  • the electronic device 9 of this embodiment includes: a processor 90 , a memory 91 , and computer-readable instructions 92 stored in the memory 91 and executable on the processor 90 .
  • the processor 90 executes the computer-readable instructions 92
  • the steps in each of the foregoing embodiments of the vehicle driving environment abnormality monitoring method are implemented, for example, steps S101 to S103 shown in FIG. 1 .
  • the processor 90 executes the computer-readable instructions 92
  • the functions of the modules/units in each of the foregoing apparatus embodiments for example, the functions of the units 81 to 83 shown in FIG. 8 , are implemented.
  • the computer-readable instructions 92 may be divided into one or more modules/units, and the one or more modules/units are stored in the memory 91 and executed by the processor 90, to complete this application.
  • the one or more modules/units may be a series of computer-readable instruction segments capable of accomplishing specific functions, and the instruction segments are used to describe the execution process of the computer-readable instructions 92 in the electronic device 9 .
  • the electronic device 9 may be a vehicle-mounted intelligent terminal.
  • the electronic device 9 may include, but is not limited to, a processor 90 and a memory 91 .
  • FIG. 9 is only an example of the electronic device 9, and does not constitute a limitation on the electronic device 9. It may include more or less components than the one shown, or combine some components, or different components
  • the electronic device 9 may also include an input and output device, a network access device, a bus, and the like.
  • the processor 90 may be a central processing unit (Central Processing Unit, CPU), and may also be other general-purpose processors, digital signal processors (Digital Signal Processors, DSP), application specific integrated circuits (Application Specific Integrated Circuits) Integrated Circuit, ASIC), off-the-shelf programmable gate array (Field-Programmable Gate Array, FPGA) or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc.
  • a general purpose processor may be a microprocessor or the processor may be any conventional processor or the like.
  • the memory 91 may be an internal storage unit of the electronic device 9 , such as a hard disk or a memory of the electronic device 9 .
  • the memory 91 may also be an external storage device of the electronic device 9, such as a plug-in hard disk, a smart memory card (Smart Media Card, SMC), a secure digital (Secure Digital, SD) equipped on the electronic device 9 card, Flash Card, etc.
  • the memory 91 may also include both an internal storage unit of the electronic device 9 and an external storage device.
  • the memory 91 is used to store the computer readable instructions and other programs and data required by the electronic device.
  • the memory 91 can also be used to temporarily store data that has been output or will be output.
  • the integrated unit if implemented in the form of a software functional unit and sold or used as an independent product, may be stored in a computer-readable storage medium.
  • the present application implements all or part of the processes in the methods of the above embodiments, which can be completed by instructing relevant hardware through computer-readable instructions, and the computer-readable instructions can be stored in a computer-readable storage medium.
  • the computer-readable instructions when executed by the processor, can implement the steps of each of the foregoing method embodiments.
  • the computer-readable instructions include computer-readable instruction codes
  • the computer-readable instruction codes may be in source code form, object code form, executable file, or some intermediate form, and the like.
  • the computer-readable medium may include at least: any entity or device capable of carrying computer-readable instruction codes to the device/terminal device, recording medium, computer memory, read-only memory (ROM, Read-Only Memory), random access Memory (RAM, Random Access Memory), electrical carrier signal, telecommunication signal and software distribution medium.
  • ROM read-only memory
  • RAM random access Memory
  • electrical carrier signal telecommunication signal and software distribution medium.
  • U disk mobile hard disk, disk or CD, etc.
  • computer readable media may not be electrical carrier signals and telecommunications signals.

Landscapes

  • Engineering & Computer Science (AREA)
  • Theoretical Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • General Physics & Mathematics (AREA)
  • Health & Medical Sciences (AREA)
  • General Health & Medical Sciences (AREA)
  • Data Mining & Analysis (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Artificial Intelligence (AREA)
  • General Engineering & Computer Science (AREA)
  • Evolutionary Computation (AREA)
  • Biomedical Technology (AREA)
  • Molecular Biology (AREA)
  • Computing Systems (AREA)
  • Computational Linguistics (AREA)
  • Biophysics (AREA)
  • Mathematical Physics (AREA)
  • Software Systems (AREA)
  • Multimedia (AREA)
  • Oral & Maxillofacial Surgery (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Human Computer Interaction (AREA)
  • Bioinformatics & Computational Biology (AREA)
  • Evolutionary Biology (AREA)
  • Bioinformatics & Cheminformatics (AREA)
  • Probability & Statistics with Applications (AREA)
  • Traffic Control Systems (AREA)

Abstract

本申请适用于安全监控技术领域，提供了一种车辆驾驶环境异常监测方法、装置、电子设备和存储介质，所述方法包括：实时获取车内图像；将所述车内图像输入至已训练完成的行为检测模型进行处理，获得所述行为检测模型输出的异常检测结果；所述行为检测模型对所述车内图像的处理包括：获取所述车内图像中人员的关键点特征及其关联嵌入值；根据所述关键点特征及其关联嵌入值，确定所述人员的行为状态；基于所述行为状态，确定所述车内是否存在异常行为；若根据所述异常检测结果确定所述车内存在异常行为，则将所述异常行为上报至指定终端。本申请可实现行车过程中对车内司机和乘客的行为的监督，从而有效保证司机和乘客的安全。

Description

车辆驾驶环境异常监测方法、装置、电子设备和存储介质 技术领域
本申请涉及安全监测技术领域，尤其涉及一种车辆驾驶环境异常监测方法、装置、电子设备和存储介质。
 背景技术
随着我国社会主义现代化建设的快速发展，城市人口逐渐增多，各种类型的交通工具涌进了人民的日常生活，交通工具增多一方面极大地方便了人民的出行，然而另一方面也带来了大量的安全问题，尤其是在人们日常生活中发挥的巨大作用的公共交通工具，诸如，网约车、出租车、公交车等。近年来，公交司机被乘客殴打导致发生重大交通事故的事件经常发生，乘客殴打司机这一现象也经常在出租车上演，网约车上乘客安全受到危险的事件时有发生。如何保障乘客和司机的安全成为我们研究的首要问题。
目前，各公共交通运营平台通常会提供一些投诉或报警的通道，利用这些投诉或报警的通道对司机的行为进行监督，但通过投诉或报警的方式通常是在异常行为发生之后进行，时效性差，并且，对于乘客骚扰伤害司机的异常行为很难及时通知运营平台或警方，行车过程中司机和乘客的安全无法得到保障。
综上所述，现有技术中，在车辆驾驶环境中缺乏对异常行为的有效监测，无法有效保障行车过程中司机和乘客安全。
 技术问题
本申请实施例提供了一种车辆驾驶环境异常监测方法、装置、电子设备和存储介质，可以解决现有技术中存在的在车辆驾驶环境中缺乏对异常行为的有效监测，无法有效保障行车过程中司机和乘客安全的问题。
 技术解决方案
为解决上述技术问题，本申请实施例采用的技术方案是：
第一方面，本申请实施例提供了一种车辆驾驶环境异常监测方法，包括：
实时获取车内图像；
将所述车内图像输入至已训练完成的行为检测模型进行处理，获得所述行为检测模型输出的异常检测结果；
其中，所述行为检测模型对所述车内图像的处理，包括：获取所述车内图像中人员的关键点特征及其关联嵌入值，所述关联嵌入值用于标识关键点特征之间的关联程度；根据所述关键点特征及其关联嵌入值，确定所述人员的行为状态；基于所述行为状态，确定所述车内是否存在异常行为；
若根据所述异常检测结果确定所述车内存在异常行为，则将所述异常行为上报至指定终端。
在第一方面的一种可能的实现方式中，所述行为检测模型包括特征提取模块、注意力模块和特征融合模块，所述行为检测模型还包括双头解耦结构；
所述获取所述车内图像中人员的关键点特征及其关联嵌入值的步骤，包括：
将所述车内图像输入至所述特征提取模块，通过所述特征提取模块输出所述人员的人体特征；
将所述人体特征输入至所述注意力模块，通过所述注意力模块对所述人体特征进行自适应加权；
将所述注意力模块的输出结果输入至所述特征融合模块，通过所述特征融合模块对所述注意力模块的输出结果进行特征融合；
利用所述双头解耦结构对所述特征融合模块的输出结果进行任务回归，得到所述人员的关键点特征及其关联嵌入值。
在第一方面的一种可能的实现方式中，所述特征提取模块包括第一特征提取子模块、第二特征提取子模块以及第三特征提取子模块，所述注意力模块包括通道注意力子模块和空间注意力子模块；
所述通过所述特征提取模块输出所述人员的人体特征的步骤，包括：
通过所述第一特征提取子模块输出第一层级的第一人体特征；
通过所述第二特征提取子模块输出第二层级的第二人体特征；
通过所述第三特征提取子模块输出第三层级的第三人体特征，其中，所述第一层级、所述第二层级以及所述第三层级存在递进关系；
所述将所述人体特征输入至所述注意力模块，通过所述注意力模块对所述人体特征进行自适应加权的步骤包括：
通过所述通道注意力子模块对所述第一人体特征、所述第二人体特征以及所述第三人体特征进行通道自适应加权；
通过所述空间注意力子模块对所述第一人体特征、所述第二人体特征以及所述第三人体特征进行空间自适应加权；
所述通过所述特征融合模块对所述注意力模块的输出结果进行特征融合的步骤，包括：
通过所述特征融合模块将所述通道注意力子模块的输出结果与所述空间注意力子模块的输出结果进行特征融合。
在第一方面的一种可能的实现方式中，所述行为检测模型还包括时空图卷积模块，所述根据所述关键点特征及其关联嵌入值，确定所述人员的行为状态的步骤，包括：
根据所述关键点特征与所述关联嵌入值，得到目标特征；
将所述目标特征输入至时空图卷积模块进行处理，获取所述时空图卷积模块输出的所述人员的行为状态；其中，所述时空图卷积模块用于对所述目标特征进行空间维度的第一图卷积与时间维度的第二图卷积，根据所述第一图卷积的结果与所述第二图卷积的结果确定所述人员的行为状态。
在第一方面的一种可能的实现方式中，所述车辆驾驶环境异常监测方法还包括：
根据预设算法，对所述已训练完成的行为检测模型进行模型优化，得到目标行为检测模型；
利用所述目标行为检测模型对所述车内图像进行处理，获得所述目标行为检测模型输出的异常检测结果；
若根据所述目标行为检测模型输出的异常检测结果确定所述车内存在异常行为，则将所述异常行为上报至所述指定终端。
在第一方面的一种可能的实现方式中，所述根据预设算法，对所述已训练完成的行为检测模型进行模型优化，得到目标行为检测模型的步骤，包括：
获取目标训练样本集；
获取教师网络模型和学生网络模型，其中，所述教师网络模型为所述已训练完成的行为检测模型，所述学生网络模型为按预设参数进行剪枝后的所述已训练完成的行为检测模型；
基于对抗生成网络与所述训练样本集，对所述教师网络模型和所述学生网络模型进行模型蒸馏，得到目标行为检测模型。
第二方面，本申请实施例提供了一种车辆驾驶环境异常监测装置，包括：
图像获取单元，用于实时获取车内图像；
异常检测单元，用于将所述车内图像输入至已训练完成的行为检测模型进行处理，获得所述行为检测模型输出的异常检测结果；其中，所述行为检测模型对所述车内图像的处理，包括：获取所述车内图像中人员的关键点特征及其关联嵌入值，所述关联嵌入值用于标识关键点特征之间的关联程度；根据所述关键点特征及其关联嵌入值，确定所述人员的行为状态；基于所述行为状态，确定所述车内是否存在异常行为；
异常上报单元，用于若根据所述异常检测结果确定所述车内存在异常行为，则将所述异常行为上报至指定终端。
第三方面，本申请实施例提供了一种电子设备，包括存储器、处理器以及存储在所述存储器中并可在所述处理器上运行的计算机可读指令，所述处理器执行所述计算机可读指令时实现如上述第一方面所述的车辆驾驶环境异常监测方法。
第四方面，本申请实施例提供了一种计算机可读存储介质，所述计算机可读存储介质存储有计算机可读指令，所述计算机可读指令被处理器执行时实现如上述第一方面所述的车辆驾驶环境异常监测方法。
第五方面，本申请实施例提供了一种计算机可读指令产品，当计算机可读指令产品在电子设备上运行时，使电子设备执行如上述第一方面所述的车辆驾驶环境异常监测方法。
 有益效果
本申请实施例中，通过实时获取车内图像，将所述车内图像输入至已训练完成的行为检测模型进行处理，获得所述行为检测模型输出的异常检测结果，其中，所述行为检测模型对所述车内图像的处理，包括：获取所述车内图像中人员的关键点特征及其关联嵌入值，所述关联嵌入值用于标识关键点特征之间的关联程度；根据所述关键点特征及其关联嵌入值，确定所述人员的行为状态；基于所述行为状态，确定所述车内是否存在异常行为，利用行为检测模型对车内人员的行为进行准确有效的异常检测，若根据所述异常检测结果确定所述车内存在异常行为，则将所述异常行为上报至指定终端，实现行车过程中对车内司机和乘客的行为的监督，以便运营平台或者相关执法单位可及时采取措施，避免意外的发生，从而有效保证司机和乘客的安全。
 附图说明
为了更清楚地说明本申请实施例中的技术方案，下面将对实施例或现有技术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申请的一些实施例，对于本领域普通技术人员来讲，在不付出创造性劳动性的前提下，还可以根据这些附图获得其他的附图。
图1是本申请实施例提供的车辆驾驶环境异常监测方法的实现流程图；
图2是本申请实施例提供的车辆驾驶环境异常监测方法中获取关键点特征及其关联嵌入值的具体实现流程图；
图3是本申请实施例提供的车辆驾驶环境异常监测方法中通过特征提取模块提取人体特征的一种具体实现流程图；
图4是本申请实施例提供的车辆驾驶环境异常监测方法中通过注意力模块进行自适应加权的一种具体实现流程图；
图5是本申请实施例提供的车辆驾驶环境异常监测方法中确定行为状态的一种具体实现流程图；
图6是本申请实施例提供的车辆驾驶环境异常监测方法中领域子集的分配策略的示意图；
图7是本申请实施例提供的车辆驾驶环境异常监测方法中模型优化的一种具体实现流程图；
图8是本申请实施例提供的车辆驾驶环境异常监测装置的结构框图；
图9是本申请实施例提供的电子设备的示意图。
 本发明的实施方式
以下描述中，为了说明而不是为了限定，提出了诸如特定系统结构、技术之类的具体细节，以便透彻理解本申请实施例。然而，本领域的技术人员应当清楚，在没有这些具体细节的其它实施例中也可以实现本申请。在其它情况中，省略对众所周知的系统、装置、电路以及方法的详细说明，以免不必要的细节妨碍本申请的描述。
应当理解，当在本申请说明书和所附权利要求书中使用时，术语“包括”指示所描述特征、整体、步骤、操作、元素和/或组件的存在，但并不排除一个或多个其它特征、整体、步骤、操作、元素、组件和/或其集合的存在或添加。
还应当理解，在本申请说明书和所附权利要求书中使用的术语“和/或”是指相关联列出的项中的一个或多个的任何组合以及所有可能组合，并且包括这些组合。
如在本申请说明书和所附权利要求书中所使用的那样，术语“如果”可以依据上下文被解释为“当...时”或“一旦”或“响应于确定”或“响应于检测到”。类似地，短语“如果确定”或“如果检测到[所描述条件或事件]”可以依据上下文被解释为意指“一旦确定”或“响应于确定”或“一旦检测到[所描述条件或事件]”或“响应于检测到[所描述条件或事件]”。
另外，在本申请说明书和所附权利要求书的描述中，术语“第一”、“第二”、“第三”等仅用于区分描述，而不能理解为指示或暗示相对重要性。
在本申请说明书中描述的参考“一个实施例”或“一些实施例”等意味着在本申请的一个或多个实施例中包括结合该实施例描述的特定特征、结构或特点。由此，在本说明书中的不同之处出现的语句“在一个实施例中”、“在一些实施例中”、“在其他一些实施例中”、“在另外一些实施例中”等不是必然都参考相同的实施例，而是意味着“一个或多个但不是所有的实施例”，除非是以其他方式另外特别强调。术语“包括”、“包含”、“具有”及它们的变形都意味着“包括但不限于”，除非是以其他方式另外特别强调。
本申请实施例提供的车辆驾驶环境异常监测方法可应用于车载智能终端。本申请不仅适用于出租车、公交车等营业车，也适用于顺风车、亲友便车、校车等情境。本申请实施例对终端设备的具体类型不作任何限制。
图1示出了本申请实施例提供的车辆驾驶环境异常监测方法的实现流程，本申请实施例执行端为车载智能终端，该方法流程包括步骤S101至S103。各步骤的具体实现原理如下：
S101：实时获取车内图像。
本申请实施例中，车载智能终端设有摄像头，利用该车载智能终端的摄像头实时拍摄车内图像。上述车内图像为待处理图像。
在一些实施方式中，实时获取车内视频，上述车内视频由一系列帧视频图像组成，从所述车内视频中提取出若干帧的车内图像。
在一种可能的实施方式中，从车载智能终端的摄像头拍摄的车内视频中抽取指定帧数的视频图像，再根据预设的图像选择算法，从抽取的指定帧数的视频图像中选择若干帧视频图像作为待处理的车内图像，帧数可由用户自定义。
示例性地，从车载智能终端的摄像头拍摄的车内视频中抽取18帧视频图像，再根据预设的图像抽取算法，从抽取的多帧视频图像中选择10帧车内图像作为待处理的车内图像。
实际上，从车载智能终端的摄像头拍摄的车内视频中抽取指定帧数的车内图像的过程，也是对车内图像进行初筛选的过程，筛选的标准可以根据图像的清晰程度、像中是否包含人体特征等确定。
在一些实施方式中，对车载智能终端的摄像头拍摄的多帧视频图像进行人脸特征点检测，将未包含人脸的视频图像筛除，得到包含人脸的视频图像，再根据预设的抽取算法，从包含人脸的视频图像中抽取指定帧数的视频图像，利用预设的图像选择算法，对所述指定帧数的视频图像进行质量择优，选择质量相对较佳的视频图像作为待处理的车内图像。
本申请实施例中的质量择优，是使用质量算法判断图像的质量，然后输出一个相应的数值，通过各图像对应的数值的比较结果，确定质量最佳的图像为待处理的车内图像。
由于车载智能终端的摄像头捕获的是视频，如果对视频内每一帧视频图像都进行处理，不仅会加大计算量，还会造成大量的冗余。在本申请实施例中，通过预设的图像选择算法进行质量择优，可减少计算量，避免冗余，选择质量最佳的视频图像作为待处理的车内图像，有利于后续处理的准确性，进而提高人脸检测处理的有效性。
上述预设的抽取算法可以是随机抽取算法，从上述经过初步筛选的视频图像中随机抽取设定数量的视频图像。
在一些实施方式中，还可以是根据预设图像标准对视频图像进行筛选，抽取指定帧数的满足所述预设图像标准的视频图像作为待处理的车内图像，预设图像标准包括图像清晰度、人脸完整度、人体关键部位个数或者人脸角度等。
S102：将所述车内图像输入至已训练完成的行为检测模型进行处理，获得所述行为检测模型输出的异常检测结果。
上述行为检测模型为深度学习神经网络模型，可以采用各类已知行为类别的样本图像作为训练样本集训练得到，比如可以采用Kinetics行为数据集中的样本图像。
其中，为了提高异常行为检测的准确性，上述行为检测模型对所述车内图像的处理，包括：获取所述车内图像中人员的关键点特征及其关联嵌入值，所述关联嵌入值用于标识关键点特征之间的关联程度，根据所述关键点特征及其关联嵌入值，确定所述人员的行为状态，基于所述行为状态，确定所述车内是否存在异常行为。
在本申请实施例中，上述已训练完成的行为检测模型的输入为若干帧的所述车内图像。上述行为检测模型要获取各帧车内图像中人员的关键点特征，同时，获取关键点特征对应的关联嵌入值，利用该关联嵌入值即可确定各关键点特征之间的关联程度。
在一些实施方式中，上述关联嵌入值的距离越近，表示关键点特征之间的关联程度越高；反之，上述关联嵌入值的距离越远，表示关键点特征之间的关联程度越低。
在一种可能的实施方式中，根据上述车内图像中人员的关键点特征及其关联嵌入值，确定上述车内图像中人员的个数。
在本申请实施例中，当不同关键点特征属于同一人员时，该不同关键点特征分别对应的关联嵌入值之间的距离较近，当不同关键点特征属于不同人员时，该不同关键点特征分别对应的关联嵌入值之间的距离较远。
在一些实施方式中，通过计算各关键点特征对应的关联嵌入值之间的差值，根据该差值确定相似关键点特征，将相似关键点特征归为同一特征群组，根据特征群组的数量确定人员的数量。
一种实施方式中，将关联嵌入值之间的差值小于或等于预设差值阈值的关键点特征确定为相似关键点特征。
示例性地，利用匈牙利算法将关联嵌入值相似的关键点特征归为同一特征组。对于匈牙利算法的介绍和使用参见现有技术，此处不赘述。
作为本申请一种可能的实施方式，上述行为检测模型包括特征提取模块、注意力模块和特征融合模块，图2示出了上述获取所述车内图像中人员的关键点特征及其关联嵌入值的具体实现流程：
A1：将所述车内图像输入至所述特征提取模块，通过所述特征提取模块输出所述人员的人体特征。
在本申请实施例中，上述特征提取模块用于提取并输出上述车内图像中人员的人体特征。人体特征包括人体骨骼特征。
在一些实施方式中，上述特征提取模块包括Moblienet v3网络，利用Moblienet v3网络输出得到司机与乘客的人体骨骼特征。
作为本申请一种可能的实施方式，上述特征提取模块包括第一特征提取子模块、第二特征提取子模块以及第三特征提取子模块，如图3所示，上述通过上述特征提取模块输出上述人员的人体特征的步骤，包括：
A11：通过上述第一特征提取子模块输出第一层级的第一人体特征。上述第一特征提取子模块用于提取第一层级的特征。
A12：通过上述第二特征提取子模块输出第二层级的第二人体特征。上述第二特征提取子模块用于提取第二层级的特征。
A13：通过上述第三特征提取子模块输出第三层级的第三人体特征，其中，上述第一层级、上述第二层级以及上述第三层级存在递进关系。上述第三特征提取子模块用于提取第三层级的特征。
本申请实施例中，第一特征提取子模块、第二特征提取子模块以及第三特征提取子模块分别从上述车内图像提取不同层级的特征。通过获取不同层级的特征丰富特征信息，有利于提高行为检测模型检测的准确性。
在一实施方式中，上述第一层级表示浅层语义，上述第二层级表示中层语义，上述第三层级表示深层语义。
在另一实施方式中，上述第一层级、上述第二层级以及上述第三层级存在递减关系，上述第一层级表示深层语义，上述第二层级表示中层语义，上述第三层级表示浅层语义。
在一实施方式中，上述第一层级表示第一分辨率，上述第一特征提取子模块用于提取第一分辨率的人体特征；上述第二层级表示第一分辨率，上述第二特征提取子模块用于提取第二分辨率的人体特征；上述第三层级表示第三分辨率，上述第三特征提取子模块用于提取第三分辨率的人体特征。其中，上述第一分辨率低于上述第二分辨率，上述第二分辨率低于上述第三分辨率。
A2：将所述人体特征输入至所述注意力模块，通过所述注意力模块对所述人体特征进行自适应加权。
在本申请实施例中，上述注意力模块利用注意力机制控制加权系数，根据该加权系数对上述特征提取模块提取的人体特征进行自适应加权。上述注意力模块的输出结果包括经过自适应加权的人体特征。
作为本申请一种可能的实施方式，上述注意力模块包括通道注意力子模块和空间注意力子模块。如图4所示，上述将上述人体特征输入至上述注意力模块，通过上述注意力模块对上述人体特征进行自适应加权的步骤包括：
A21：通过上述通道注意力子模块对上述第一人体特征、上述第二人体特征以及上述第三人体特征进行通道自适应加权。
A22：通过所述空间注意力子模块对所述第一人体特征、所述第二人体特征以及所述第三人体特征进行空间自适应加权。
在本申请实施例中，利用通道注意力子模块和空间注意力子模块分别对第一人体特征、第二人体特征以及第三人体特征进行挤压和激励，实现对第一人体特征、第二人体特征以及第三人体特征通道和空间的自适应加权。
在一种可能的实施方式中，将上述第一人体特征、上述第二人体特征以及上述第三人体特征进行特征拼接后输入至上述注意力模块。
A3：将所述注意力模块的输出结果输入至所述特征融合模块，通过所述特征融合模块对所述注意力模块的输出结果进行特征融合。
上述特征融合模块用于将自适应加权后的人体特征进行融合，以利于进一步分析处理。
一些实施方式中，通过上述特征融合模块将上述通道注意力子模块的输出结果与上述空间注意力子模块的输出结果进行特征融合。
A4：利用所述双头解耦结构对所述特征融合模块的输出结果进行任务回归，得到所述人员的关键点特征及其关联嵌入值。
在本申请实施例中，上述行为检测模型还包括双头解耦结构。耦合是指两个或两个以上的体系或两种运动形式间通过相互作用而彼此影响以至联合起来的现象。解耦就是用数学方法将两种运动分离开来处理问题。模块间有依赖关系必然存在耦合。
在一种可能的实施方式中，上述特征融合模块的输出结果包括关键点热图和关联嵌入值，利用上述双头解耦结构进行任务回归。任务回归是指确定关键点特征所属人员的标签。其中，上述关键点热图采用损失函数focal loss进行回归，关联嵌入值则使用损失函数push loss配合损失函数pull loss实现回归。
本申请实施例利用双头解耦结构降低模块之间的耦合度，避免信息回传耦合严重。
示例性地，利用Moblienet v3网络的C2、C3、C4三个子特征提模块分别提取浅层、中层以及深层的人体特征，然后利用通道注意力子模块和空间注意力子模块分别对C2、C3、C4提取的人体特征进行通道和空间的自适应加权。利用特征融合模块将自适应加权后的人体特征进行整合，再利用双头解耦结构对特征融合模块的输出结果进行任务回归，得到上述车内图像中人员的关键点特征及其关联嵌入值。
作为本申请一种可能的实施方式，所述行为检测模型还包括时空图卷积模块，图5示出了上述根据所述关键点特征及其关联嵌入值，确定所述人员的行为状态的具体实现流程：
B1：根据所述关键点特征与所述关联嵌入值，得到目标特征。
本申请实施例中，上述目标特征为上述关键点特征与上述关联嵌入值进行特征向量拼接得到的。
B2：将所述目标特征输入至时空图卷积模块进行处理，获取所述时空图卷积模块输出的所述人员的行为状态。其中，所述时空图卷积模块用于对所述目标特征进行空间维度的第一图卷积与时间维度的第二图卷积，根据所述第一图卷积的结果与所述第二图卷积的结果确定所述人员的行为状态。
本申请实施例中，上述时空图卷积模块包括三层时空图卷积。首先，定义领域子集的分配策略，如图6所示，根据领域内节点子集到人体中心的距离划分为3个子集，分别为靠近人体中心的第一子集，远离人体中心的第二子集和自身节点的第三子集。然后，根据领域子集的定义，时空图卷积就可以类似与普通卷积一样，先进行空间维度的第一图卷积，即不同子集乘以不同的关键点特征向量，同一子集内乘以相同的关键点特征向量，再进行时间维度的第二图卷积，通过简单的三层时空图卷积叠加后，利用softmax函数对车内人员的行为状态进行回归。通过这样的方式，既有效地利用了人体骨骼关键点的结构先验，也有效的利用了时间维度上的信息传递，从而可准确判断车内人员的行为状态，提高车内驾驶环境检异常监测的准确性。
S103：若根据所述异常检测结果确定所述车内存在异常行为，则将所述异常行为上报至指定终端。
本申请实施例中，上述人员包括司机和乘客。上述行为状态分为正常行为状态和异常行为状态。上述行为检测模型输出的异常检测结果包括人员的行为状态，若所述行为状态为异常行为状态，则上述异常检测结果还包括异常行为。异常行为包括乘客骚扰、殴打司机，司机骚扰、殴打乘客。
本申请实施例中，当确定车内存在异常行为时，立即根据该异常行为生成通报事件上传至指定终端。上述车载智能终端与指定终端通信连接。上述指定终端包括但不限于运营平台的智能终端、监管部门的智能终端，以及指定用户的移动终端。
在一些可能的实施方式中，上述车辆驾驶环境异常监测方法还包括：若根据所述异常检测结果确定所述车内存在异常行为，则发送声音警报，通过该声音警报提示车内司机和乘客停止上述异常行为。
作为本申请一种可能的实施方式，车辆驾驶环境异常监测方法还包括：
C1：根据预设算法，对所述已训练完成的行为检测模型进行模型优化，得到目标行为检测模型。
本申请实施例中，为了进一步适应边缘计算设备的算力需求，对上述已训练完成的行为检测模型进行模型优化。
在一种可能的实施方式中，上述预设算法包括模型蒸馏，如图7所示，上述步骤C1包括：
C11：获取目标训练样本集。该目标训练样本集中的样本数及样本类型可以根据实际需要确定，本申请在此不做具体限定。也可以直接采用上述行为检测模型的训练样本集。
C12：获取教师网络模型和学生网络模型，其中，所述教师网络模型为所述已训练完成的行为检测模型，所述学生网络模型为按预设参数进行剪枝后的所述已训练完成的行为检测模型。上述预设参数可以为模型敏感度。
上述剪枝是指剪掉模型中卷积层的通道数。例如，本申请实施例中，剪枝率设置为0.4，剪掉上述已训练完成的行为检测模型中单个卷积层40%的通道。
C13：基于对抗生成网络与所述训练样本集，对所述教师网络模型和所述学生网络模型进行模型蒸馏，得到目标行为检测模型。
本申请实施例中，在进行模型蒸馏的过程中加入对抗生成网络的思想，可有效保证模型的精度。
C2：利用所述目标行为检测模型对所述车内图像进行处理，获得所述目标行为检测模型输出的异常检测结果。
在本申请实施例中，将已训练完成的行为检测模型确定为原始模型，并作为教师网络，将基于敏感度分析自动剪枝后的已训练完成的行为检测模作为学生网络，进行模型蒸馏。蒸馏的过程中加入GAN的思想，进行模型蒸馏时我们将教师网络作为判别模型，学生网络作为生成模型，用判别网络有效的监督生成网络进行模型的迭代优化，最终得到算法精度高，计算量低的网络模型。
C3：若根据所述目标行为检测模型输出的异常检测结果确定所述车内存在异常行为，则将所述异常行为上报至所述指定终端。
本申请实施例中，使用结构化剪枝配合模型蒸馏对模型进一步进行压缩，从而得到优化后的目标行为检测模型，利用目标行为检测模型对所述车内图像进行处理，获得所述目标行为检测模型输出的异常检测结果，在避免模型精度损失的同时，大大减少模型计算量，可增强异常行为监测的实时性。
由上可见，本申请实施例中，通过实时获取车内图像，将所述车内图像输入至已训练完成的行为检测模型进行处理，获得所述行为检测模型输出的异常检测结果，其中，所述行为检测模型对所述车内图像的处理，包括：获取所述车内图像中人员的关键点特征及其关联嵌入值，所述关联嵌入值用于标识关键点特征之间的关联程度；根据所述关键点特征及其关联嵌入值，确定所述人员的行为状态；基于所述行为状态，确定所述车内是否存在异常行为，利用行为检测模型对车内人员的行为进行准确有效的异常检测，若根据所述异常检测结果确定所述车内存在异常行为，则将所述异常行为上报至指定终端，实现行车过程中对车内司机和乘客的行为的监督，以便运营平台或者相关执法单位可及时采取措施，避免意外的发生，从而有效保证司机和乘客的安全。
应理解，上述实施例中各步骤的序号的大小并不意味着执行顺序的先后，各过程的执行顺序应以其功能和内在逻辑确定，而不应对本申请实施例的实施过程构成任何限定。
对应于上文实施例所述的车辆驾驶环境异常监测方法，图8示出了本申请实施例提供的车辆驾驶环境异常监测装置的结构框图，为了便于说明，仅示出了与本申请实施例相关的部分。
参照图8，该车辆驾驶环境异常监测装置包括：图像获取单元81，异常检测单元82，异常上报单元83，其中：
图像获取单元81，用于实时获取车内图像；
异常检测单元82，用于将所述车内图像输入至已训练完成的行为检测模型进行处理，获得所述行为检测模型输出的异常检测结果；其中，所述行为检测模型对所述车内图像的处理，包括：获取所述车内图像中人员的关键点特征及其关联嵌入值，所述关联嵌入值用于标识关键点特征之间的关联程度；根据所述关键点特征及其关联嵌入值，确定所述人员的行为状态；基于所述行为状态，确定所述车内是否存在异常行为；
异常上报单元83，用于若根据所述异常检测结果确定所述车内存在异常行为，则将所述异常行为上报至指定终端。
作为本申请一种可能的实施方式，上述行为检测模型包括特征提取模块、注意力模块和特征融合模块，上述行为检测模型还包括双头解耦结构；上述异常检测单元82具体用于：
将所述车内图像输入至所述特征提取模块，通过所述特征提取模块输出所述人员的人体特征；
将所述人体特征输入至所述注意力模块，通过所述注意力模块对所述人体特征进行自适应加权；
将所述注意力模块的输出结果输入至所述特征融合模块，通过所述特征融合模块对所述注意力模块的输出结果进行特征融合；
利用所述双头解耦结构对所述特征融合模块的输出结果进行任务回归，得到所述人员的关键点特征及其关联嵌入值。
作为本申请一种可能的实施方式，上述特征提取模块包括第一特征提取子模块、第二特征提取子模块以及第三特征提取子模块，上述注意力模块包括通道注意力子模块和空间注意力子模块；
上述通过所述特征提取模块输出所述人员的人体特征的步骤，包括：
通过所述第一特征提取子模块输出第一层级的第一人体特征；
通过所述第二特征提取子模块输出第二层级的第二人体特征；
通过所述第三特征提取子模块输出第三层级的第三人体特征，其中，所述第一层级、所述第二层级以及所述第三层级存在递进关系；
上述将所述人体特征输入至所述注意力模块，通过所述注意力模块对所述人体特征进行自适应加权的步骤包括：
通过所述通道注意力子模块对所述第一人体特征、所述第二人体特征以及所述第三人体特征进行通道自适应加权；
通过所述空间注意力子模块对所述第一人体特征、所述第二人体特征以及所述第三人体特征进行空间自适应加权；
上述通过所述特征融合模块对所述注意力模块的输出结果进行特征融合的步骤，包括：
通过所述特征融合模块将所述通道注意力子模块的输出结果与所述空间注意力子模块的输出结果进行特征融合。
作为本申请一种可能的实施方式，上述行为检测模型还包括时空图卷积模块，上述异常检测单元82还用于：
根据所述关键点特征与所述关联嵌入值，得到目标特征；
将所述目标特征输入至时空图卷积模块进行处理，获取所述时空图卷积模块输出的所述人员的行为状态；其中，所述时空图卷积模块用于对所述目标特征进行空间维度的第一图卷积与时间维度的第二图卷积，根据所述第一图卷积的结果与所述第二图卷积的结果确定所述人员的行为状态。
作为本申请一种可能的实施方式，所述车辆驾驶环境异常监测装置还包括：
模型优化单元，用于根据预设算法，对所述已训练完成的行为检测模型进行模型优化，得到目标行为检测模型；
上述异常检测单元82，还用于利用所述目标行为检测模型对所述车内图像进行处理，获得所述目标行为检测模型输出的异常检测结果；
上述异常上报单元83，还用于若根据所述目标行为检测模型输出的异常检测结果确定所述车内存在异常行为，则将所述异常行为上报至所述指定终端。
作为本申请一种可能的实施方式，上述模型优化单元包括：
样本获取模块，用于获取目标训练样本集；
优化训练模块，用于获取教师网络模型和学生网络模型，其中，所述教师网络模型为所述已训练完成的行为检测模型，所述学生网络模型为按预设参数进行剪枝后的所述已训练完成的行为检测模型；
目标模型生成模块，用于基于对抗生成网络与所述训练样本集，对所述教师网络模型和所述学生网络模型进行模型蒸馏，得到目标行为检测模型。
由上可见，本申请实施例中，通过实时获取车内图像，将所述车内图像输入至已训练完成的行为检测模型进行处理，获得所述行为检测模型输出的异常检测结果，其中，所述行为检测模型对所述车内图像的处理，包括：获取所述车内图像中人员的关键点特征及其关联嵌入值，所述关联嵌入值用于标识关键点特征之间的关联程度；根据所述关键点特征及其关联嵌入值，确定所述人员的行为状态；基于所述行为状态，确定所述车内是否存在异常行为，利用行为检测模型对车内人员的行为进行准确有效的异常检测，若根据所述异常检测结果确定所述车内存在异常行为，则将所述异常行为上报至指定终端，实现行车过程中对车内司机和乘客的行为的监督，以便运营平台或者相关执法单位可及时采取措施，避免意外的发生，从而有效保证司机和乘客的安全。
需要说明的是，上述装置/单元之间的信息交互、执行过程等内容，由于与本申请方法实施例基于同一构思，其具体功能及带来的技术效果，具体可参见方法实施例部分，此处不再赘述。
本申请实施例还提供一种计算机可读存储介质，所述计算机可读存储介质存储有计算机可读指令，所述计算机可读指令被处理器执行时实现如图1至图7表示的任意一种车辆驾驶环境异常监测方法的步骤。
本申请实施例还提供一种电子设备，包括存储器、处理器以及存储在所述存储器中并可在所述处理器上运行的计算机可读指令，所述处理器执行所述计算机可读指令时实现如图1至图7表示的任意一种车辆驾驶环境异常监测方法的步骤。
本申请实施例还提供一种计算机可读指令产品，当该计算机可读指令产品在电子设备上运行时，使得电子设备执行实现如图1至图7表示的任意一种车辆驾驶环境异常监测方法的步骤。
图9是本申请一实施例提供的电子设备的示意图。如图9所示，该实施例的电子设备9包括：处理器90、存储器91以及存储在所述存储器91中并可在所述处理器90上运行的计算机可读指令92。所述处理器90执行所述计算机可读指令92时实现上述各个车辆驾驶环境异常监测方法实施例中的步骤，例如图1所示的步骤S101至S103。或者，所述处理器90执行所述计算机可读指令92时实现上述各装置实施例中各模块/单元的功能，例如图8所示单元81至83的功能。
示例性的，所述计算机可读指令92可以被分割成一个或多个模块/单元，所述一个或者多个模块/单元被存储在所述存储器91中，并由所述处理器90执行，以完成本申请。所述一个或多个模块/单元可以是能够完成特定功能的一系列计算机可读指令段，该指令段用于描述所述计算机可读指令92在所述电子设备9中的执行过程。
所述电子设备9可以为车载智能终端。所述电子设备9可包括，但不仅限于，处理器90、存储器91。本领域技术人员可以理解，图9仅仅是电子设备9的示例，并不构成对电子设备9的限定，可以包括比图示更多或更少的部件，或者组合某些部件，或者不同的部件，例如所述电子设备9还可以包括输入输出设备、网络接入设备、总线等。
所述处理器90可以是中央处理单元(Central Processing Unit，CPU)，还可以是其他通用处理器、数字信号处理器 (Digital Signal Processor，DSP)、专用集成电路 (Application Specific Integrated Circuit，ASIC)、现成可编程门阵列 (Field-Programmable Gate Array，FPGA) 或者其他可编程逻辑器件、分立门或者晶体管逻辑器件、分立硬件组件等。通用处理器可以是微处理器或者该处理器也可以是任何常规的处理器等。
所述存储器91可以是所述电子设备9的内部存储单元，例如电子设备9的硬盘或内存。所述存储器91也可以是所述电子设备9的外部存储设备，例如所述电子设备9上配备的插接式硬盘，智能存储卡（Smart Media Card, SMC），安全数字（Secure Digital, SD）卡，闪存卡（Flash Card）等。进一步地，所述存储器91还可以既包括所述电子设备9的内部存储单元也包括外部存储设备。所述存储器91用于存储所述计算机可读指令以及所述电子设备所需的其他程序和数据。所述存储器91还可以用于暂时地存储已经输出或者将要输出的数据。
需要说明的是，上述装置/单元之间的信息交互、执行过程等内容，由于与本申请方法实施例基于同一构思，其具体功能及带来的技术效果，具体可参见方法实施例部分，此处不再赘述。
所属领域的技术人员可以清楚地了解到，为了描述的方便和简洁，仅以上述各功能单元、模块的划分进行举例说明，实际应用中，可以根据需要而将上述功能分配由不同的功能单元、模块完成，即将所述装置的内部结构划分成不同的功能单元或模块，以完成以上描述的全部或者部分功能。实施例中的各功能单元、模块可以集成在一个处理单元中，也可以是各个单元单独物理存在，也可以两个或两个以上单元集成在一个单元中，上述集成的单元既可以采用硬件的形式实现，也可以采用软件功能单元的形式实现。另外，各功能单元、模块的具体名称也只是为了便于相互区分，并不用于限制本申请的保护范围。上述系统中单元、模块的具体工作过程，可以参考前述方法实施例中的对应过程，在此不再赘述。
所述集成的单元如果以软件功能单元的形式实现并作为独立的产品销售或使用时，可以存储在一个计算机可读取存储介质中。基于这样的理解，本申请实现上述实施例方法中的全部或部分流程，可以通过计算机可读指令来指令相关的硬件来完成，所述的计算机可读指令可存储于一计算机可读存储介质中，该计算机可读指令在被处理器执行时，可实现上述各个方法实施例的步骤。其中，所述计算机可读指令包括计算机可读指令代码，所述计算机可读指令代码可以为源代码形式、对象代码形式、可执行文件或某些中间形式等。所述计算机可读介质至少可以包括：能够将计算机可读指令代码携带到装置/终端设备的任何实体或装置、记录介质、计算机存储器、只读存储器（ROM，Read-Only Memory）、随机存取存储器（RAM，Random Access Memory）、电载波信号、电信信号以及软件分发介质。例如U盘、移动硬盘、磁碟或者光盘等。在某些司法管辖区，根据立法和专利实践，计算机可读介质不可以是电载波信号和电信信号。
在上述实施例中，对各个实施例的描述都各有侧重，某个实施例中没有详述或记载的部分，可以参见其它实施例的相关描述。
以上所述实施例仅用以说明本申请的技术方案，而非对其限制；尽管参照前述实施例对本申请进行了详细的说明，本领域的普通技术人员应当理解：其依然可以对前述各实施例所记载的技术方案进行修改，或者对其中部分技术特征进行等同替换；而这些修改或者替换，并不使相应技术方案的本质脱离本申请各实施例技术方案的精神和范围，均应包含在本申请的保护范围之内。

Claims (20)

  1. 一种车辆驾驶环境异常监测方法，其特征在于，包括：
    实时获取车内图像；
    将所述车内图像输入至已训练完成的行为检测模型进行处理，获得所述行为检测模型输出的异常检测结果；
    其中，所述行为检测模型对所述车内图像的处理，包括：获取所述车内图像中人员的关键点特征及其关联嵌入值，所述关联嵌入值用于标识关键点特征之间的关联程度；根据所述关键点特征及其关联嵌入值，确定所述人员的行为状态；基于所述行为状态，确定所述车内是否存在异常行为；
    若根据所述异常检测结果确定所述车内存在异常行为，则将所述异常行为上报至指定终端。
  2. 根据权利要求1所述的车辆驾驶环境异常监测方法，其特征在于，所述行为检测模型包括特征提取模块、注意力模块和特征融合模块，所述行为检测模型还包括双头解耦结构；
    所述获取所述车内图像中人员的关键点特征及其关联嵌入值的步骤，包括：
    将所述车内图像输入至所述特征提取模块，通过所述特征提取模块输出所述人员的人体特征；
    将所述人体特征输入至所述注意力模块，通过所述注意力模块对所述人体特征进行自适应加权；
    将所述注意力模块的输出结果输入至所述特征融合模块，通过所述特征融合模块对所述注意力模块的输出结果进行特征融合；
    利用所述双头解耦结构对所述特征融合模块的输出结果进行任务回归，得到所述人员的关键点特征及其关联嵌入值。
  3. 根据权利要求2所述的车辆驾驶环境异常监测方法，其特征在于，所述特征提取模块包括第一特征提取子模块、第二特征提取子模块以及第三特征提取子模块，所述注意力模块包括通道注意力子模块和空间注意力子模块；
    所述通过所述特征提取模块输出所述人员的人体特征的步骤，包括：
    通过所述第一特征提取子模块输出第一层级的第一人体特征；
    通过所述第二特征提取子模块输出第二层级的第二人体特征；
    通过所述第三特征提取子模块输出第三层级的第三人体特征，其中，所述第一层级、所述第二层级以及所述第三层级存在递进关系；
    所述将所述人体特征输入至所述注意力模块，通过所述注意力模块对所述人体特征进行自适应加权的步骤包括：
    通过所述通道注意力子模块对所述第一人体特征、所述第二人体特征以及所述第三人体特征进行通道自适应加权；
    通过所述空间注意力子模块对所述第一人体特征、所述第二人体特征以及所述第三人体特征进行空间自适应加权；
    所述通过所述特征融合模块对所述注意力模块的输出结果进行特征融合的步骤，包括：
    通过所述特征融合模块将所述通道注意力子模块的输出结果与所述空间注意力子模块的输出结果进行特征融合。
  4. 根据权利要求1所述的车辆驾驶环境异常监测方法，其特征在于，所述行为检测模型还包括时空图卷积模块，所述根据所述关键点特征及其关联嵌入值，确定所述人员的行为状态的步骤，包括：
    根据所述关键点特征与所述关联嵌入值，得到目标特征；
    将所述目标特征输入至时空图卷积模块进行处理，获取所述时空图卷积模块输出的所述人员的行为状态；其中，所述时空图卷积模块用于对所述目标特征进行空间维度的第一图卷积与时间维度的第二图卷积，根据所述第一图卷积的结果与所述第二图卷积的结果确定所述人员的行为状态。
  5. 根据权利要求1至4任一项所述的车辆驾驶环境异常监测方法，其特征在于，所述车辆驾驶环境异常监测方法还包括：
    根据预设算法，对所述已训练完成的行为检测模型进行模型优化，得到目标行为检测模型；
    利用所述目标行为检测模型对所述车内图像进行处理，获得所述目标行为检测模型输出的异常检测结果；
    若根据所述目标行为检测模型输出的异常检测结果确定所述车内存在异常行为，则将所述异常行为上报至所述指定终端。
  6. 根据权利要求5所述的车辆驾驶环境异常监测方法，其特征在于，所述根据预设算法，对所述已训练完成的行为检测模型进行模型优化，得到目标行为检测模型的步骤，包括：
    获取目标训练样本集；
    获取教师网络模型和学生网络模型，其中，所述教师网络模型为所述已训练完成的行为检测模型，所述学生网络模型为按预设参数进行剪枝后的所述已训练完成的行为检测模型；
    基于对抗生成网络与所述训练样本集，对所述教师网络模型和所述学生网络模型进行模型蒸馏，得到目标行为检测模型。
  7. 一种车辆驾驶环境异常监测装置，其特征在于，包括：
    图像获取单元，用于实时获取车内图像；
    异常检测单元，用于将所述车内图像输入至已训练完成的行为检测模型进行处理，获得所述行为检测模型输出的异常检测结果；其中，所述行为检测模型对所述车内图像的处理，包括：获取所述车内图像中人员的关键点特征及其关联嵌入值，所述关联嵌入值用于标识关键点特征之间的关联程度；根据所述关键点特征及其关联嵌入值，确定所述人员的行为状态；基于所述行为状态，确定所述车内是否存在异常行为；
    异常上报单元，用于若根据所述异常检测结果确定所述车内存在异常行为，则将所述异常行为上报至指定终端。
  8. 根据权利要求7所述的车辆驾驶环境异常监测装置，其特征在于，所述行为检测模型包括特征提取模块、注意力模块和特征融合模块，所述行为检测模型还包括双头解耦结构；所述异常检测单元具体用于：
    将所述车内图像输入至所述特征提取模块，通过所述特征提取模块输出所述人员的人体特征；
    将所述人体特征输入至所述注意力模块，通过所述注意力模块对所述人体特征进行自适应加权；
    将所述注意力模块的输出结果输入至所述特征融合模块，通过所述特征融合模块对所述注意力模块的输出结果进行特征融合；
    利用所述双头解耦结构对所述特征融合模块的输出结果进行任务回归，得到所述人员的关键点特征及其关联嵌入值。
  9. 根据权利要求8所述的车辆驾驶环境异常监测装置，其特征在于，所述特征提取模块包括第一特征提取子模块、第二特征提取子模块以及第三特征提取子模块，所述注意力模块包括通道注意力子模块和空间注意力子模块；
    所述通过所述特征提取模块输出所述人员的人体特征的步骤，包括：
    通过所述第一特征提取子模块输出第一层级的第一人体特征；
    通过所述第二特征提取子模块输出第二层级的第二人体特征；
    通过所述第三特征提取子模块输出第三层级的第三人体特征，其中，所述第一层级、所述第二层级以及所述第三层级存在递进关系；
    所述将所述人体特征输入至所述注意力模块，通过所述注意力模块对所述人体特征进行自适应加权的步骤包括：
    通过所述通道注意力子模块对所述第一人体特征、所述第二人体特征以及所述第三人体特征进行通道自适应加权；
    通过所述空间注意力子模块对所述第一人体特征、所述第二人体特征以及所述第三人体特征进行空间自适应加权；
    所述通过所述特征融合模块对所述注意力模块的输出结果进行特征融合的步骤，包括：
    通过所述特征融合模块将所述通道注意力子模块的输出结果与所述空间注意力子模块的输出结果进行特征融合。
  10. 根据权利要求7所述的车辆驾驶环境异常监测装置，其特征在于，所述异常检测单元还用于：
    根据所述关键点特征与所述关联嵌入值，得到目标特征；
    将所述目标特征输入至时空图卷积模块进行处理，获取所述时空图卷积模块输出的所述人员的行为状态；其中，所述时空图卷积模块用于对所述目标特征进行空间维度的第一图卷积与时间维度的第二图卷积，根据所述第一图卷积的结果与所述第二图卷积的结果确定所述人员的行为状态。
  11. 根据权利要求7至10任一项所述的车辆驾驶环境异常监测装置，其特征在于，所述车辆驾驶环境异常监测装置还包括：
    模型优化单元，用于根据预设算法，对所述已训练完成的行为检测模型进行模型优化，得到目标行为检测模型；
    所述异常检测单元，还用于利用所述目标行为检测模型对所述车内图像进行处理，获得所述目标行为检测模型输出的异常检测结果；
    所述异常上报单元，还用于若根据所述目标行为检测模型输出的异常检测结果确定所述车内存在异常行为，则将所述异常行为上报至所述指定终端。
  12. 根据权利要求11所述的车辆驾驶环境异常监测装置，其特征在于，所述模型优化单元包括：
    样本获取模块，用于获取目标训练样本集；
    优化训练模块，用于获取教师网络模型和学生网络模型，其中，所述教师网络模型为所述已训练完成的行为检测模型，所述学生网络模型为按预设参数进行剪枝后的所述已训练完成的行为检测模型；
    目标模型生成模块，用于基于对抗生成网络与所述训练样本集，对所述教师网络模型和所述学生网络模型进行模型蒸馏，得到目标行为检测模型。
  13. 一种电子设备，包括存储器、处理器以及存储在所述存储器中并可在所述处理器上运行的计算机可读指令，其特征在于，所述处理器执行所述计算机可读指令时实现如下步骤：
    实时获取车内图像；
    将所述车内图像输入至已训练完成的行为检测模型进行处理，获得所述行为检测模型输出的异常检测结果；
    其中，所述行为检测模型对所述车内图像的处理，包括：获取所述车内图像中人员的关键点特征及其关联嵌入值，所述关联嵌入值用于标识关键点特征之间的关联程度；根据所述关键点特征及其关联嵌入值，确定所述人员的行为状态；基于所述行为状态，确定所述车内是否存在异常行为；
    若根据所述异常检测结果确定所述车内存在异常行为，则将所述异常行为上报至指定终端。
  14. 根据权利要求13所述的电子设备，其特征在于，所述行为检测模型包括特征提取模块、注意力模块和特征融合模块，所述行为检测模型还包括双头解耦结构；
    所述获取所述车内图像中人员的关键点特征及其关联嵌入值的步骤，包括：
    将所述车内图像输入至所述特征提取模块，通过所述特征提取模块输出所述人员的人体特征；
    将所述人体特征输入至所述注意力模块，通过所述注意力模块对所述人体特征进行自适应加权；
    将所述注意力模块的输出结果输入至所述特征融合模块，通过所述特征融合模块对所述注意力模块的输出结果进行特征融合；
    利用所述双头解耦结构对所述特征融合模块的输出结果进行任务回归，得到所述人员的关键点特征及其关联嵌入值。
  15. 根据权利要求13所述的电子设备，其特征在于，所述行为检测模型还包括时空图卷积模块，所述根据所述关键点特征及其关联嵌入值，确定所述人员的行为状态的步骤，包括：
    根据所述关键点特征与所述关联嵌入值，得到目标特征；
    将所述目标特征输入至时空图卷积模块进行处理，获取所述时空图卷积模块输出的所述人员的行为状态；其中，所述时空图卷积模块用于对所述目标特征进行空间维度的第一图卷积与时间维度的第二图卷积，根据所述第一图卷积的结果与所述第二图卷积的结果确定所述人员的行为状态。
  16. 根据权利要求13至15任一项所述的电子设备，其特征在于，所述处理器执行所述计算机可读指令时还实现如下步骤：
    根据预设算法，对所述已训练完成的行为检测模型进行模型优化，得到目标行为检测模型；
    利用所述目标行为检测模型对所述车内图像进行处理，获得所述目标行为检测模型输出的异常检测结果；
    若根据所述目标行为检测模型输出的异常检测结果确定所述车内存在异常行为，则将所述异常行为上报至所述指定终端。
  17. 一种计算机可读存储介质，所述计算机可读存储介质存储有计算机可读指令，其特征在于，所述计算机可读指令被处理器执行时实现如下步骤：
    实时获取车内图像；
    将所述车内图像输入至已训练完成的行为检测模型进行处理，获得所述行为检测模型输出的异常检测结果；
    其中，所述行为检测模型对所述车内图像的处理，包括：获取所述车内图像中人员的关键点特征及其关联嵌入值，所述关联嵌入值用于标识关键点特征之间的关联程度；根据所述关键点特征及其关联嵌入值，确定所述人员的行为状态；基于所述行为状态，确定所述车内是否存在异常行为；
    若根据所述异常检测结果确定所述车内存在异常行为，则将所述异常行为上报至指定终端。
  18. 根据权利要求17所述的计算机可读存储介质，其特征在于，所述行为检测模型包括特征提取模块、注意力模块和特征融合模块，所述行为检测模型还包括双头解耦结构；
    所述获取所述车内图像中人员的关键点特征及其关联嵌入值的步骤，包括：
    将所述车内图像输入至所述特征提取模块，通过所述特征提取模块输出所述人员的人体特征；
    将所述人体特征输入至所述注意力模块，通过所述注意力模块对所述人体特征进行自适应加权；
    将所述注意力模块的输出结果输入至所述特征融合模块，通过所述特征融合模块对所述注意力模块的输出结果进行特征融合；
    利用所述双头解耦结构对所述特征融合模块的输出结果进行任务回归，得到所述人员的关键点特征及其关联嵌入值。
  19. 根据权利要求17所述的计算机可读存储介质，其特征在于，所述行为检测模型还包括时空图卷积模块，所述根据所述关键点特征及其关联嵌入值，确定所述人员的行为状态的步骤，包括：
    根据所述关键点特征与所述关联嵌入值，得到目标特征；
    将所述目标特征输入至时空图卷积模块进行处理，获取所述时空图卷积模块输出的所述人员的行为状态；其中，所述时空图卷积模块用于对所述目标特征进行空间维度的第一图卷积与时间维度的第二图卷积，根据所述第一图卷积的结果与所述第二图卷积的结果确定所述人员的行为状态。
  20. 根据权利要求17至19任一项所述的计算机可读存储介质，其特征在于，所述计算机可读指令被处理器执行时还实现如下步骤：
    根据预设算法，对所述已训练完成的行为检测模型进行模型优化，得到目标行为检测模型；
    利用所述目标行为检测模型对所述车内图像进行处理，获得所述目标行为检测模型输出的异常检测结果；
    若根据所述目标行为检测模型输出的异常检测结果确定所述车内存在异常行为，则将所述异常行为上报至所述指定终端。
PCT/CN2021/086050 2021-04-09 2021-04-09 车辆驾驶环境异常监测方法、装置、电子设备和存储介质 WO2022213336A1 (zh)

Priority Applications (2)

Application Number Priority Date Filing Date Title
PCT/CN2021/086050 WO2022213336A1 (zh) 2021-04-09 2021-04-09 车辆驾驶环境异常监测方法、装置、电子设备和存储介质
CN202180000757.9A CN113287120A (zh) 2021-04-09 2021-04-09 车辆驾驶环境异常监测方法、装置、电子设备和存储介质

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
PCT/CN2021/086050 WO2022213336A1 (zh) 2021-04-09 2021-04-09 车辆驾驶环境异常监测方法、装置、电子设备和存储介质

Publications (1)

Publication Number Publication Date
WO2022213336A1 true WO2022213336A1 (zh) 2022-10-13

Family

ID=77281369

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2021/086050 WO2022213336A1 (zh) 2021-04-09 2021-04-09 车辆驾驶环境异常监测方法、装置、电子设备和存储介质

Country Status (2)

Country Link
CN (1) CN113287120A (zh)
WO (1) WO2022213336A1 (zh)

Cited By (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN115829584A (zh) * 2022-12-02 2023-03-21 首约科技(北京)有限公司 飘点的确定方法、装置、电子设备及存储介质
CN117152964A (zh) * 2023-11-01 2023-12-01 宁波宁工交通工程设计咨询有限公司 一种基于行驶车辆的城市道路信息智能采集方法

Families Citing this family (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN114005105B (zh) * 2021-12-30 2022-04-12 青岛以萨数据技术有限公司 驾驶行为检测方法、装置以及电子设备
CN114612762A (zh) * 2022-03-15 2022-06-10 首约科技(北京)有限公司 一种智能设备监管方法
CN116704666A (zh) * 2023-06-21 2023-09-05 合肥中科类脑智能技术有限公司 售卖方法及计算机可读存储介质、自动售卖机

Citations (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN109522793A (zh) * 2018-10-10 2019-03-26 华南理工大学 基于机器视觉的多人异常行为检测与识别方法
CN111027478A (zh) * 2019-12-10 2020-04-17 青岛农业大学 一种基于深度学习的司机、乘客行为分析与预警系统
CN111681454A (zh) * 2020-06-03 2020-09-18 重庆邮电大学 一种基于驾驶行为的车车协同防撞预警方法
CN111860254A (zh) * 2020-07-10 2020-10-30 东莞正扬电子机械有限公司 一种驾驶员异常行为检测方法、装置、存储介质及设备
CN112613441A (zh) * 2020-12-29 2021-04-06 新疆爱华盈通信息技术有限公司 异常驾驶行为的识别和预警方法、电子设备

Patent Citations (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN109522793A (zh) * 2018-10-10 2019-03-26 华南理工大学 基于机器视觉的多人异常行为检测与识别方法
CN111027478A (zh) * 2019-12-10 2020-04-17 青岛农业大学 一种基于深度学习的司机、乘客行为分析与预警系统
CN111681454A (zh) * 2020-06-03 2020-09-18 重庆邮电大学 一种基于驾驶行为的车车协同防撞预警方法
CN111860254A (zh) * 2020-07-10 2020-10-30 东莞正扬电子机械有限公司 一种驾驶员异常行为检测方法、装置、存储介质及设备
CN112613441A (zh) * 2020-12-29 2021-04-06 新疆爱华盈通信息技术有限公司 异常驾驶行为的识别和预警方法、电子设备

Cited By (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN115829584A (zh) * 2022-12-02 2023-03-21 首约科技(北京)有限公司 飘点的确定方法、装置、电子设备及存储介质
CN117152964A (zh) * 2023-11-01 2023-12-01 宁波宁工交通工程设计咨询有限公司 一种基于行驶车辆的城市道路信息智能采集方法
CN117152964B (zh) * 2023-11-01 2024-02-02 宁波宁工交通工程设计咨询有限公司 一种基于行驶车辆的城市道路信息智能采集方法

Also Published As

Publication number Publication date
CN113287120A (zh) 2021-08-20

Similar Documents

Publication Publication Date Title
WO2022213336A1 (zh) 车辆驾驶环境异常监测方法、装置、电子设备和存储介质
CN110414313B (zh) 异常行为告警方法、装置、服务器及存储介质
CN109523652B (zh) 基于驾驶行为的保险的处理方法、装置、设备及存储介质
CN111325319B (zh) 一种神经网络模型的检测方法、装置、设备及存储介质
CN111881707B (zh) 图像翻拍检测方法、身份验证方法、模型训练方法及装置
CN112200081A (zh) 异常行为识别方法、装置、电子设备及存储介质
US20210125005A1 (en) System and method for protection and detection of adversarial attacks against a classifier
CN114373189A (zh) 一种行为检测方法、装置、终端设备及存储介质
CN116186770A (zh) 图像脱敏方法、装置、电子设备及存储介质
CN113839904A (zh) 基于智能网联汽车的安全态势感知方法和系统
CN114170585B (zh) 危险驾驶行为的识别方法、装置、电子设备及存储介质
CN111444788B (zh) 行为识别的方法、装置及计算机存储介质
Zhan et al. A structural variation classification model for image quality assessment
CN111563468A (zh) 一种基于神经网络注意力的驾驶员异常行为检测方法
CN113486856A (zh) 一种基于语义分割和卷积神经网络的驾驶员不规范行为检测方法
CN111565303B (zh) 基于雾计算及深度学习的视频监控方法、系统和可读存储介质
CN116939164A (zh) 一种安防监控方法及系统
CN110909641A (zh) 一种检测摩托车超载的方法、装置及系统
CN115731401A (zh) 一种基于纹理感知的图像烟雾精细检测方法
CN115861980A (zh) 一种集成多种特征信号的驾驶疲劳检测方法、装置与系统
CN115713751A (zh) 疲劳驾驶检测方法、设备、存储介质及装置
CN109410582B (zh) 交通状况分析方法及终端设备
CN113537087A (zh) 一种智慧交通信息处理方法、装置及服务器
CN115357646B (zh) 一种桥梁状态监测方法及系统
CN113807428B (zh) 分类模型概率标签的重构方法、系统、装置及存储介质

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: 21935566

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: 21935566

Country of ref document: EP

Kind code of ref document: A1