WO2020173056A1 - 交通图像识别方法、装置、计算机设备和介质 - Google Patents

交通图像识别方法、装置、计算机设备和介质 Download PDF

Info

Publication number
WO2020173056A1
WO2020173056A1 PCT/CN2019/102027 CN2019102027W WO2020173056A1 WO 2020173056 A1 WO2020173056 A1 WO 2020173056A1 CN 2019102027 W CN2019102027 W CN 2019102027W WO 2020173056 A1 WO2020173056 A1 WO 2020173056A1
Authority
WO
WIPO (PCT)
Prior art keywords
interference
picture
types
autoencoder
traffic
Prior art date
Application number
PCT/CN2019/102027
Other languages
English (en)
French (fr)
Inventor
刘焱
王洋
郝新
吴月升
Original Assignee
百度在线网络技术（北京）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 百度在线网络技术（北京）有限公司 filed Critical 百度在线网络技术（北京）有限公司
Priority to EP19916553.1A priority Critical patent/EP3786835A4/en
Priority to JP2020568528A priority patent/JP2022521448A/ja
Priority to KR1020207035694A priority patent/KR20210031427A/ko
Publication of WO2020173056A1 publication Critical patent/WO2020173056A1/zh
Priority to US17/114,076 priority patent/US20210117705A1/en

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V10/00—Arrangements for image or video recognition or understanding
    • G06V10/20—Image preprocessing
    • G06V10/30—Noise filtering
    • 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
    • G06V20/56—Context or environment of the image exterior to a vehicle by using sensors mounted on the vehicle
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
    • G06F18/214—Generating training patterns; Bootstrap methods, e.g. bagging or boosting
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N20/00—Machine learning
    • 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
    • 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
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T3/00—Geometric image transformations in the plane of the image
    • G06T3/14—Transformations for image registration, e.g. adjusting or mapping for alignment of images
    • G06T3/147—Transformations for image registration, e.g. adjusting or mapping for alignment of images using affine transformations
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T5/00—Image enhancement or restoration
    • G06T5/70—Denoising; Smoothing
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V10/00—Arrangements for image or video recognition or understanding
    • G06V10/20—Image preprocessing
    • G06V10/26—Segmentation of patterns in the image field; Cutting or merging of image elements to establish the pattern region, e.g. clustering-based techniques; Detection of occlusion
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V10/00—Arrangements for image or video recognition or understanding
    • G06V10/70—Arrangements for image or video recognition or understanding using pattern recognition or machine learning
    • G06V10/764—Arrangements for image or video recognition or understanding using pattern recognition or machine learning using classification, e.g. of video objects
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V10/00—Arrangements for image or video recognition or understanding
    • G06V10/70—Arrangements for image or video recognition or understanding using pattern recognition or machine learning
    • G06V10/77—Processing image or video features in feature spaces; using data integration or data reduction, e.g. principal component analysis [PCA] or independent component analysis [ICA] or self-organising maps [SOM]; Blind source separation
    • G06V10/774—Generating sets of training patterns; Bootstrap methods, e.g. bagging or boosting
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V10/00—Arrangements for image or video recognition or understanding
    • G06V10/70—Arrangements for image or video recognition or understanding using pattern recognition or machine learning
    • G06V10/82—Arrangements for image or video recognition or understanding using pattern recognition or machine learning using neural networks
    • 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
    • G06V20/56—Context or environment of the image exterior to a vehicle by using sensors mounted on the vehicle
    • G06V20/58—Recognition of moving objects or obstacles, e.g. vehicles or pedestrians; Recognition of traffic objects, e.g. traffic signs, traffic lights or roads
    • G06V20/582—Recognition of moving objects or obstacles, e.g. vehicles or pedestrians; Recognition of traffic objects, e.g. traffic signs, traffic lights or roads of traffic signs

Definitions

  • the embodiments of the present application relate to the technical field of automatic driving image processing, for example, to a traffic image recognition method, device, computer equipment, and medium.
  • unmanned vehicles will obtain traffic lights, traffic signs and other information in the form of video streams.
  • the driving control system preprocesses the video obtained by the camera or radar to obtain a picture with characteristic information, and then inputs the picture with characteristic information into the classification model of traffic lights and traffic signs for prediction, such as judging whether it is a red light or a green light , Is the speed limit of 60 kilometers or a stop sign.
  • the classification model in the unmanned vehicle system is usually a deep learning model, which is very easy to be attacked by adversarial samples and cause misjudgment. For example, sticking a small picture sticker on a street sign or traffic indicator, and constructing an adversarial sample on the small picture to make the classification model misjudge, it will not be able to recognize the street sign or indicator normally, which will affect the driving safety of unmanned vehicles. .
  • the embodiments of the present application provide a traffic image recognition method, device, computer equipment, and medium to reduce the interference of counter-samples in traffic images, improve the accuracy of image recognition, and improve the safety of intelligent driving.
  • an embodiment of the present application provides a traffic image recognition method.
  • the method includes: acquiring a video stream collected by a vehicle and extracting each frame of image in the video stream as a first picture; Input to the anti-interference autoencoder for preprocessing, to filter the interference in the first picture, and output the second picture, where the anti-interference autoencoder is obtained by training at least two types of interference sample sets, different types
  • the disturbance modes added to the interference sample set include at least two of the following: noise, affine change, filter blurring, brightness change, and monochromatic; the second picture is input to the traffic sign recognition model for recognition processing.
  • an embodiment of the present application also provides a traffic image recognition device.
  • the device includes: a picture acquisition module configured to acquire a video stream collected by a vehicle and extract each frame of image in the video stream as a first picture ; Picture preprocessing module, configured to input the first picture to the interference-removing self-encoder for preprocessing, so as to filter the interference in the first picture, and output a second picture, wherein the interference-removing self-encoder It is obtained by training at least two types of interference sample sets.
  • the interference types added to different types of interference sample sets include at least two of the following: noise, affine change, filter blurring, brightness change, and monochromatic; image recognition module, It is configured to input the second picture into a traffic sign recognition model for recognition processing.
  • an embodiment of the present application also provides a computer device, which includes: one or more processors; a storage device configured to store at least one program; when the at least one program is processed by the at least one program The device executes, so that the at least one processor implements the traffic image recognition method described in any of the embodiments of the present application.
  • an embodiment of the present application also provides a computer-readable storage medium on which a computer program is stored, and when the program is executed by a processor, the traffic image recognition method as described in any of the embodiments of the present application is implemented.
  • the pictures in the video stream collected by the vehicle are input to the interference-removing self-encoder, and the interference-removed pictures are obtained through the preprocessing of the interference-removing self-encoder, and then the images without interference are input to the traffic sign recognition model
  • Perform recognition processing to facilitate subsequent generation of correct vehicle control instructions, and solve the problem of traffic sign recognition errors caused by adversarial samples attacking the traffic sign recognition model; it can reduce the interference of adversarial samples in traffic images and improve the accuracy of image recognition. Improve the safety of driverless or intelligent driving.
  • Fig. 1 is a flow chart of the traffic image recognition method in the first embodiment of the present application
  • Figure 2a is a flowchart of a traffic image recognition method in Embodiment 2 of the present application.
  • 2b is a schematic diagram of the neural network structure of the autoencoder in the second embodiment of the present application.
  • FIG. 3 is a schematic structural diagram of a traffic image recognition device in Embodiment 3 of the present application.
  • FIG. 4 is a schematic diagram of the structure of the computer equipment in the fourth embodiment of the present application.
  • Figure 1 is a flow chart of the traffic image recognition method provided by the first embodiment. This embodiment can be applied to resist attacks based on countermeasure samples on the road signs and traffic light recognition models of unmanned vehicles or intelligent driving control systems.
  • the method may be implemented by a traffic image recognition device, and specifically may be implemented by software and/or hardware in the device, for example, an unmanned vehicle or a vehicle driving control system in a smartly-driving vehicle.
  • the traffic image recognition method specifically includes:
  • S110 Obtain a video stream collected by the vehicle and extract each frame of image in the video stream as a first picture.
  • the vehicle can be an unmanned vehicle or a vehicle with intelligent driving function.
  • the above two types of vehicles are equipped with cameras, radars or cameras and radars, which are used to collect the moving direction of the vehicle and the video stream around the vehicle.
  • the image content in the video stream usually includes traffic signs, signal lights, lane lines, and other vehicles, pedestrians, and buildings.
  • the collected video stream will be transmitted to the control system of the vehicle, and then the control system will extract each frame of image, the first picture, from the video stream as the target object for analysis.
  • Each frame of image extracted above can be understood as a target image that is determined to be traffic sign recognition after other processing.
  • the perturbation methods added in different types of interference sample sets include at least two of the following: noise, affine change, filter blurring, brightness change, and monochromatic.
  • the first picture it may or may not contain traffic signs, signal lights, or lane lines that have the function of traffic indication.
  • the first picture including the traffic instruction information usually plays a decisive role in the control of the vehicle.
  • traffic signs such as traffic signs, signal lights or lane lines will be interfered with by means of advertisements, labels or overlays, etc., so that the traffic sign recognition model cannot correctly identify the traffic signs, resulting in violation of traffic rules , Even endangering the personal safety of passengers and the safety of public transportation.
  • the first picture can be input to the anti-interference autoencoder for preprocessing, and the interference information can be filtered out when there is interference information in the first image containing traffic sign information, and the second picture, that is, an image without interference .
  • the preprocessing of the anti-interference autoencoder does not have a significant impact on the picture, and an output image close to the original image can be obtained.
  • the anti-interference autoencoder is obtained by training at least two types of interference sample sets. It can not only filter out the interference caused by the interference processing of a single picture, but also filter out the interference caused by the combination of multiple interference processing methods, and improve the anti-sample image. Disturbance filtering effect.
  • each type of anti-interference sample set there is at least one sample pair, and each sample pair includes an original picture and a confrontation sample corresponding to the original picture.
  • each adversarial sample undergoes the same type of disturbance processing relative to the corresponding original picture.
  • the so-called same type refers to the same combination of disturbance methods.
  • the combination of disturbance modes may include a single disturbance mode, or may also include a combination of two or more disturbance modes.
  • the combination of disturbance modes used is the same, but the specific parameters used in each disturbance mode can be the same or different.
  • the perturbation methods used in the embodiments of the present application may have multiple types.
  • the perturbation methods include at least two of noise, affine change, filter blurring, brightness change, and monochromatic.
  • the first picture before the first picture is input to the interference-removing self-encoder for preprocessing, the first picture can also be compressed from the color dimension, that is, RGB color information, gray scale or RGB color Compression processing in terms of information and gray levels.
  • the color dimension that is, RGB color information, gray scale or RGB color Compression processing in terms of information and gray levels.
  • the traffic sign recognition model is usually a network model based on deep learning.
  • the traffic sign recognition model can identify the characteristic information in the second image and determine whether the characteristic information belongs to any traffic sign, such as a speed limit sign or a traffic light, for the decision-making module of the vehicle driving control system to recognize according to the traffic sign recognition model As a result, a control decision is made to control the vehicle during driving.
  • any traffic sign such as a speed limit sign or a traffic light
  • the pictures in the video stream collected by the vehicle are input to the anti-interference autoencoder, and the interference-free pictures are obtained through the pre-processing of the anti-interference autoencoder, and then the interference-free pictures are input to the traffic
  • the sign recognition model performs recognition processing, so that the subsequent correct vehicle control instructions can be generated, which solves the problem of traffic sign recognition errors caused by the adversarial samples attacking the traffic sign recognition model; it can reduce the interference of the adversarial samples in the traffic image and improve the image quality
  • the recognition accuracy rate improves the safety of driverless or intelligent driving.
  • the technical solutions of the embodiments of the present application can be applied to both black box attacks initiated by illegal users when they are uncertain of the deep learning model used for traffic sign recognition, and white box attacks when the deep learning model is determined.
  • Black box attacks are different from white box attacks.
  • White box attacks are often targeted to use the Fast Gradient Sign Method (FGSM), CW (Clarke and Wright), and Jacobian-based saliency mapping when the model structure and specific parameters of the deep learning model are known.
  • FGSM Fast Gradient Sign Method
  • CW Clarke and Wright
  • Jacobian-based saliency mapping when the model structure and specific parameters of the deep learning model are known.
  • Methods Jacobian-based saliency map approach, JSMA
  • other adversarial sample algorithms carry out white box attacks.
  • the black box attack When the black box attack is uncertain about the deep learning model, it will launch complex and changeable black box attacks through disturbances such as noise, affine changes, filter blurring, brightness changes, and monochromaticization.
  • the embodiments of the present application effectively solve the black box attack and the white box attack and filter out each disturbance, so that the deep learning model of traffic sign recognition can effectively perform recognition and filtering.
  • Fig. 2a is a flowchart of a traffic image recognition method provided in Embodiment 2 of this application. This embodiment is based on each of the optional solutions in the foregoing embodiment, and provides a training process of the anti-interference autoencoder. As shown in Figure 2a, the traffic image recognition method provided in the embodiment of the present application includes the following steps:
  • the original picture is a picture without added interference
  • the content of the picture is real traffic lights, traffic signs, lane lines and road signs.
  • the way to obtain the original picture can be through a terminal with a camera function, or it can be intercepted in a certain video.
  • the sample set is generated.
  • the original picture is processed to form an interference picture.
  • the original picture and the interference picture are used as a sample pair, and at least two types of sample pair sets are selected as the interference sample set. For each type of interference sample set, the same combination of disturbance methods is determined.
  • an affine change and a filter blurring change are added to the first original picture to generate a first interference picture, and the first original picture and the first interference picture are a sample pair.
  • the original picture is processed by one or more perturbation methods of adding noise, adding affine changes, superimposing filter blurring changes, superimposing brightness changes, and superimposing monochromatic changes.
  • at least one perturbation parameter value in any type of perturbation method can be adjusted to form at least two perturbations, thereby increasing the number of interference pictures generated for the same original picture, thereby increasing the number of sample pair sets.
  • adjusting the value of at least one disturbance parameter in any type of disturbance mode to form at least two disturbances may include at least one of the following:
  • multiple parameter values can be changed at the same time to form different interference pictures, such as changing the flip angle parameter and the clipping angle parameter in the radiation change at the same time, and the brightness value in the brightness change .
  • Auto Encoders is a common model in deep learning. Its structure is a three-layer neural network structure. The structure includes an input layer, a hidden layer and an output layer. Among them, the output layer and the input layer Have the same dimensionality, please refer to Figure 2b for details. Specifically, the input layer and output layer represent the input layer and output layer of the neural network, respectively.
  • the hidden layer undertakes the work of the encoder and decoder.
  • the encoding process is to transform from the high-dimensional input layer to the low-dimensional hidden layer.
  • the process, on the contrary, the decoding process is the conversion process from the low-dimensional hidden layer to the high-dimensional output layer. Therefore, the autoencoder is a lossy conversion process.
  • the loss function is defined by comparing the difference between input and output.
  • the training process does not need to label the data, the whole process is the process of continuously solving the loss function minimization.
  • the noise superimposed interference picture in any sample pair is input to the input layer, and then the picture restored by the hidden layer of the autoencoder is obtained in the output layer, and then the original picture and the restored picture
  • the picture is input into the loss function at the same time, and the autoencoder needs to be optimized according to the output result of the loss function.
  • the training process can be stopped, and the anti-interference autoencoder is finally obtained.
  • the interfering autoencoder can be LSTM (Long Short-Term Memory).
  • the convolutional neural network model the samples of the interference sample set include at least two consecutive pictures. That is, the original picture is an original sample group composed of at least two consecutive frame pictures, and the interference picture group corresponding to the original sample group is a picture with interference information superimposed on the original sample group in the same disturbance manner.
  • the same disturbance mode means that the same combination of disturbance modes is adopted.
  • the combination of disturbance modes may include a single disturbance mode, or may also include a combination of two or more disturbance modes.
  • the combination of disturbance modes used is the same, but the specific parameters used by each disturbance mode can be the same or different.
  • the perturbation methods used in the embodiments of the present application may have multiple types.
  • the perturbation methods include at least two of noise, affine change, filter blurring, brightness change, and monochromatic.
  • the sample images in the sample set can also be compressed from the color dimension, that is, RGB color information, gray level or RGB color information and gray level.
  • the compression processing This is because the recognition of traffic signs mainly depends on the structure, shape, and main color of the object, and is not sensitive to the color of the details. After the image is compressed in the color dimension, the amount of data in the image processing can be reduced.
  • S230 Obtain a video stream collected by the vehicle and extract each frame of image in the video stream as a first picture.
  • different types of interference sample sets are formed by adding interference noise to the original picture through different perturbation methods, and the autoencoder is trained to obtain a de-interference autoencoder that can filter out multiple interferences, and then use this
  • the anti-interference autoencoder performs anti-interference pre-processing on the pictures in the video stream collected by the vehicle to obtain pictures that filter out the interference, and input the pre-processed pictures into the traffic sign recognition model for identification processing, thereby generating correct vehicle control instructions , Solve the problem of traffic sign recognition errors caused by adversarial samples attacking the traffic sign recognition model; it can reduce the interference of adversarial samples in traffic images, improve the accuracy of image recognition, and improve the safety of unmanned driving or intelligent driving.
  • Fig. 3 shows a schematic structural diagram of a traffic image recognition device provided in the third embodiment of the present application.
  • the embodiment of the present application may be suitable for resisting the recognition model of road signs and traffic indicator lights for unmanned vehicles or intelligent driving control systems based on adversarial samples. Of the attack situation.
  • the traffic image recognition device in the embodiment of the present application includes: a picture acquisition module 310, a picture preprocessing module 320, and a picture recognition module 330.
  • the picture acquisition module 310 is configured to acquire the video stream collected by the vehicle and extract each frame of image in the video stream as the first picture;
  • the picture preprocessing module 320 is configured to input the first picture to the interference removal
  • the self-encoder performs pre-processing to filter the interference in the first picture and output a second picture, wherein the anti-interference autoencoder is obtained by training at least two types of interference sample sets, and the different types of interference sample sets are all
  • the interference types to be added include at least two of the following: noise, affine change, filter blurring, brightness change, and monochromatic;
  • the picture recognition module 330 is configured to input the second picture into the traffic sign recognition model for recognition deal with.
  • the pictures in the video stream collected by the vehicle are input to the anti-interference autoencoder, and the interference-free pictures are obtained through the pre-processing of the anti-interference autoencoder, and then the interference-free pictures are input to the traffic
  • the sign recognition model performs recognition processing to generate correct vehicle control instructions, which solves the problem of traffic sign recognition errors caused by adversarial samples attacking the traffic sign recognition model; it can reduce the interference of adversarial samples in traffic images and improve the accuracy of image recognition. Improve the safety of driverless or intelligent driving.
  • the traffic image recognition device further includes: a sample set generation module, configured to add at least two types of interference on the basis of the original picture to form at least two types of interference sample sets; a model training module, configured to set each The sample pairs in the interference sample set are respectively used as input pictures and output pictures, and are input to the autoencoder for training.
  • a sample set generation module configured to add at least two types of interference on the basis of the original picture to form at least two types of interference sample sets
  • a model training module configured to set each The sample pairs in the interference sample set are respectively used as input pictures and output pictures, and are input to the autoencoder for training.
  • the sample set generation module is configured to: obtain the original picture; by adding noise, adding affine change, superimposing filter blurring change, superimposing brightness change, superimposing monochromatic change, one or more perturbation methods , Processing the original picture to form an interference picture; taking the original picture and the interference picture as a sample pair, and selecting at least two types of sample pair sets as the interference sample set.
  • the sample set generation module is further configured to adjust at least one disturbance parameter value in any type of disturbance mode to form at least two disturbances.
  • adjusting at least one perturbation parameter value in any type of perturbation method to form at least two perturbations includes at least one of the following: adjusting a scaling parameter in the radiation change to form perturbations with different scaling ratios ; Adjust the input parameters of the fuzzy controller in the filter blurring to form disturbances with different degrees of blur; adjust the brightness value in the brightness change to form the disturbance of different brightness; adjust the pixel value of the pixel in the monochromatic change, To form disturbances of different colors.
  • the input layer and output layer of the self-encoder have the same structure, so that the output picture has the same resolution as the original picture.
  • the traffic image recognition device further includes an image compression module, configured to perform compression processing on the first picture from the color dimension before inputting the first picture to the anti-interference autoencoder for preprocessing.
  • the anti-interference autoencoder is a convolutional neural network model of LSTM, and the interference sample set includes at least two consecutive pictures.
  • the traffic image recognition device provided by the embodiment of the present application can execute the traffic image recognition method provided by any embodiment of the present application, and has the corresponding functional modules and beneficial effects of the execution method.
  • FIG. 4 is a schematic diagram of the structure of the computer equipment in the fourth embodiment of the present application.
  • Figure 4 shows a block diagram of an exemplary computer device 412 suitable for implementing embodiments of the present application.
  • the computer device 412 shown in FIG. 4 is only an example, and should not bring any limitation to the functions and scope of use of the embodiments of the present application.
  • the computer device 412 is in the form of a general-purpose computing device.
  • the components of the computer device 412 may include, but are not limited to: one or more processors or processing units 416, a system memory 428, and a bus 418 connecting different system components (including the system memory 428 and the processing unit 416).
  • the bus 418 represents one or more of several types of bus structures, including a memory bus or a memory controller, a peripheral bus, a graphics acceleration port, a processor, or a local bus using any bus structure among multiple bus structures.
  • these architectures include but are not limited to industry standard architecture (ISA) bus, microchannel architecture (MAC) bus, enhanced ISA bus, Video Electronics Standards Association (VESA) local bus, and peripheral component interconnection ( PCI) bus.
  • ISA industry standard architecture
  • MAC microchannel architecture
  • VESA Video Electronics Standards Association
  • PCI peripheral component interconnection
  • the computer device 412 typically includes a variety of computer system readable media. These media may be any available media that can be accessed by the computer device 412, including volatile and non-volatile media, removable and non-removable media.
  • the system memory 428 may include a computer system readable medium in the form of volatile memory, such as random access memory (RAM) 430 and/or cache memory 432.
  • the computer device 412 may further include other removable/non-removable, volatile/non-volatile computer system storage media.
  • the storage system 434 may be used to read and write non-removable, non-volatile magnetic media (not shown in FIG. 4, usually referred to as a "hard drive").
  • a disk drive for reading and writing to removable non-volatile disks such as "floppy disks"
  • a removable non-volatile disk such as CD-ROM, DVD-ROM
  • other optical media read and write optical disc drives.
  • each drive may be connected to the bus 418 through one or more data media interfaces.
  • the memory 428 may include at least one program product, and the program product has a set (for example, at least one) program modules, which are configured to perform the functions of each embodiment of the present application.
  • a program/utility tool 440 having a set of (at least one) program module 442 may be stored in, for example, the memory 428.
  • Such program module 442 includes but is not limited to an operating system, one or more application programs, other program modules, and program data Each of these examples or some combination may include the implementation of a network environment.
  • the program module 442 generally executes the functions and/or methods in the embodiments described in this application.
  • the computer device 412 can also communicate with one or more external devices 414 (such as a keyboard, pointing device, display 424, etc.), and can also communicate with one or more devices that enable a user to interact with the computer device 412, and/or communicate with Any device (such as a network card, modem, etc.) that enables the computer device 412 to communicate with one or more other computing devices. This communication can be performed through an input/output (I/O) interface 422.
  • the computer device 412 may also communicate with one or more networks (for example, a local area network (LAN), a wide area network (WAN), and/or a public network, such as the Internet) through the network adapter 420.
  • networks for example, a local area network (LAN), a wide area network (WAN), and/or a public network, such as the Internet
  • the network adapter 420 communicates with other modules of the computer device 412 through the bus 418. It should be understood that although not shown in FIG. 4, other hardware and/or software modules can be used in conjunction with the computer device 412, including but not limited to: microcode, device drivers, redundant processing units, external disk drive arrays, RAID systems, tape Drives and data backup storage systems, etc.
  • the processing unit 416 executes each functional application and data processing by running the program stored in the system memory 428, for example, realizes the traffic image recognition method provided in the embodiment of the present application.
  • the method mainly includes: acquiring the video stream collected by the vehicle and Extract each frame of image in the video stream as the first picture; input the first picture to the interference-removing autoencoder for preprocessing, to filter the interference in the first picture, and output the second picture, where
  • the anti-interference autoencoder is obtained by training at least two types of interference sample sets, and the disturbance modes added to the different types of interference sample sets include at least two of the following: noise, affine change, filter blurring, brightness change, and Monochromatic; input the second picture into a traffic sign recognition model for recognition processing.
  • the fifth embodiment of the present application also provides a computer-readable storage medium on which a computer program is stored.
  • the traffic image recognition method as provided in the embodiment of the present application is implemented.
  • the method mainly includes: obtaining The video stream collected by the vehicle is extracted and each frame of image in the video stream is extracted as the first picture; the first picture is input to the anti-interference autoencoder for preprocessing to filter the interference in the first picture, Output a second picture, where the anti-interference autoencoder is obtained by training at least two types of interference sample sets, and the disturbance modes added to different types of interference sample sets include at least two of the following: noise, affine change, filtering Blur, brightness change and monochromatic; input the second picture into a traffic sign recognition model for recognition processing.
  • the computer storage media in the embodiments of the present application may adopt any combination of one or more computer-readable media.
  • the computer-readable medium may be a computer-readable signal medium or a computer-readable storage medium.
  • the computer-readable storage medium may be, for example, but not limited to, an electric, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any combination of the above.
  • computer-readable storage media include: electrical connections with one or more wires, portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), Erasable programmable read-only memory (EPROM or flash memory), optical fiber, portable compact disk read-only memory (CD-ROM), optical storage device, magnetic storage device, or any suitable combination of the above.
  • RAM random access memory
  • ROM read-only memory
  • EPROM or flash memory Erasable programmable read-only memory
  • CD-ROM compact disk read-only memory
  • the computer-readable storage medium can be any tangible medium that contains or stores a program, and the program can be used by or in combination with an instruction execution system, apparatus, or device.
  • the computer-readable signal medium may include a data signal propagated in baseband or as a part of a carrier wave, and computer-readable program code is carried therein. This propagated data signal can take many forms, including but not limited to electromagnetic signals, optical signals, or any suitable combination of the foregoing.
  • the computer-readable signal medium may also be any computer-readable medium other than the computer-readable storage medium.
  • the computer-readable medium may send, propagate, or transmit the program for use by or in combination with the instruction execution system, apparatus, or device .
  • the program code contained on the computer-readable medium can be transmitted by any suitable medium, including but not limited to wireless, wire, optical cable, RF, etc., or any suitable combination of the above.
  • the computer program code used to perform the operations of the present application can be written in one or more programming languages or a combination thereof.
  • the programming languages include object-oriented programming languages-such as Java, Smalltalk, C++, and also conventional Procedural programming language-such as "C" language or similar programming language.
  • the program code can be executed entirely on the user's computer, partly on the user's computer, executed as an independent software package, partly on the user's computer and partly executed on a remote computer, or entirely executed on the remote computer or server.
  • the remote computer can be connected to the user’s computer through any kind of network, including a local area network (LAN) or a wide area network (WAN), or it can be connected to an external computer (for example, using an Internet service provider to pass Internet connection).
  • LAN local area network
  • WAN wide area network
  • Internet service provider for example, using an Internet service provider to pass Internet connection.

Landscapes

  • Engineering & Computer Science (AREA)
  • Theoretical Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • General Physics & Mathematics (AREA)
  • Evolutionary Computation (AREA)
  • Artificial Intelligence (AREA)
  • Multimedia (AREA)
  • Software Systems (AREA)
  • Computing Systems (AREA)
  • Health & Medical Sciences (AREA)
  • General Health & Medical Sciences (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Data Mining & Analysis (AREA)
  • Medical Informatics (AREA)
  • Databases & Information Systems (AREA)
  • General Engineering & Computer Science (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Mathematical Physics (AREA)
  • Biophysics (AREA)
  • Biomedical Technology (AREA)
  • Computational Linguistics (AREA)
  • Molecular Biology (AREA)
  • Evolutionary Biology (AREA)
  • Bioinformatics & Computational Biology (AREA)
  • Bioinformatics & Cheminformatics (AREA)
  • Traffic Control Systems (AREA)
  • Image Analysis (AREA)
  • Image Processing (AREA)

Abstract

本申请实施例公开了一种交通图像识别方法、装置、设备和介质。其中，方法包括：获取车辆采集的视频流并提取出视频流中的每帧图像作为第一图片；将第一图片输入至去干扰自编码器进行预处理，以过滤第一图片中的干扰，输出第二图片，其中，去干扰自编码器通过至少两类干扰样本集进行训练而得，不同类型干扰样本集中所加入的扰动方式包括下述至少两种：噪音、仿射变化、滤波模糊化、亮度变化和单色化；将第二图片输入至交通标志识别模型来进行识别处理。本申请实施例提供了正确识别对抗样本对交通标志识别模型进行攻击后的交通标志的方式；可以降低交通图像中对抗样本的干扰，提高图像的识别正确率，提升无人驾驶或智能驾驶的安全性。

Description

交通图像识别方法、装置、计算机设备和介质
本申请要求在2019年2月25日提交中国专利局、申请号为201910138054.7的中国专利申请的优先权，该申请的全部内容通过引用结合在本申请中。
技术领域
本申请实施例涉及自动驾驶图片处理技术领域，例如涉及一种交通图像识别方法、装置、计算机设备和介质。
背景技术
无人车在行驶过程或智能驾驶控制过程中，会以视频流的形式获取交通指示灯、交通指示牌等信息。例如，驾驶控制系统对摄像头或雷达获取的视频进行预处理得到含有特征信息的图片，进而将含有特征信息图片输入交通指示灯、交通指示牌的分类模型中进行预测，比如判断是红灯还是绿灯，是限速60公里还是停车指示牌。
但是，无人车系统中的分类模型通常是深度学习模型，十分容易遭受对抗样本的攻击，产生误判。例如，在路牌或者交通指示灯上贴上小图贴，通过在小图贴上构造对抗样本，让分类模型产生误判，就不能正常识别路牌或指示灯，对无人车的驾驶安全产生影响。
发明内容
以下是对本文详细描述的主题的概述。本概述并非是为了限制权利要求的保护范围。
本申请实施例提供一种交通图像识别方法、装置、计算机设备和介质，以降低交通图像中对抗样本的干扰，提高图像的识别正确率，提升智能驾驶的安全性。
第一方面，本申请实施例提供了一种交通图像识别方法，该方法包括：获取车辆采集的视频流并提取出所述视频流中的每帧图像作为第一图片；将所述第一图片输入至去干扰自编码器进行预处理，以过滤所述第一图片中的干扰，输出第二图片，其中，所述去干扰自编码器通过至少两类干扰样本集进行训练而得，不同类型干扰样本集中所加入的扰动方式包括下述至少两种：噪音、仿射变化、滤波模糊化、亮度变化和单色化；将所述第二图片输入至交通标志识别模型来进行识别处理。
第二方面，本申请实施例还提供了一种交通图像识别装置，该装置包括：图片采集模块，设置为获取车辆采集的视频流并提取出所述视频流中的每帧图像作为第一图片；图片预处理模块，设置为将所述第一图片输入至去干扰自编码器进行预处理，以过滤所述第一图片中的干扰，输出第二图片，其中，所述去干扰自编码器通过至少两类干扰样本集进行训练而得，不同类型干扰样本集中所加入的干扰类型包括下述至少两种：噪音、仿射变化、滤波模糊化、亮度变化和单色化；图片识别模块，设置为将所述第二图片输入至交通标志识别模型来进行识别处理。
第三方面，本申请实施例还提供了一种计算机设备，该计算机设备包括：一个或多个处理器；存储装置，设置为存储至少一个程序；当所述至少一个程序被所述至少一个处理器执行，使得所述至少一个处理器实现本申请实施例中任一所述的交通图像识别方法。
第四方面，本申请实施例还提供了一种计算机可读存储介质，其上存储有计算机程序，该程序被处理器执行时实现如本申请实施例中任一所述的交通图像识别方法。
本申请实施例通过将车辆采集的视频流中的图片输入至去干扰自编码器，经过去干扰自编码器的预处理得到滤除干扰的图片，进而将无干扰的图片输入至交通标志识别模型进行识别处理，从而方便后续产生正确的车辆控制指令，解决了对抗样本对交通标志识别模型进行攻击导致交通标志识别错误的问题；可以降低交通图像中对抗样本的干扰，提高图像的识别正确率，提升无人驾驶或智能驾驶的安全性。
在阅读并理解了附图和详细描述后，可以明白其他方面。
附图说明
图1是本申请实施例一中的交通图像识别方法的流程图；
图2a是本申请实施例二中的交通图像识别方法的流程图；
图2b是本申请实施例二中的自编码器神经网络结构示意图；
图3是本申请实施例三中的交通图像识别装置的结构示意图；
图4是本申请实施例四中的计算机设备的结构示意图。
具体实施方式
下面结合附图和实施例对本申请实施例作进一步的详细说明。可以理解的是，此处所描述的具体实施例仅仅用于解释本申请实施例，而非对本申请的限定。另外还需要说明的是，为了便于描述，附图中仅示出了与本申请实施例相关的部分而非全部结构。
实施例一
图1为本实施例一提供的交通图像识别方法的流程图，本实施例可适用于抵御基于对抗样本对无人车或智能驾驶控制系统的路牌和交通指示灯识别模型的攻击的情况，该方法可以由交通图像识别装置实现，具体可通过设备中的软件和/或硬件来实施，例如，无人驾驶车辆或可智能驾驶车辆中的车辆驾驶控制系统。如图1所示，交通图像识别方法具体包括：
S110、获取车辆采集的视频流并提取出所述视频流中的每帧图像作为第一图片。
其中，车辆可以是无人驾驶的车辆，或是具有智能驾驶功能的车辆。在上述两种类型的车辆上均设置有摄像头、雷达或摄像头和雷达，用于采集车辆在行驶过程中，车辆的前进方向和辆周边的视频流。视频流中的图像内容通常包括交通标志、信号灯、车道线及其他车辆、行人和建筑物等内容。采集到的视频流会被传输到车辆的控制系统，然后，控制系统在视频流中提取出每一帧图像即第一图片，作为分析的目标对象。上述所提取出的每帧 图像可以理解为经过其他处理后确定要进行交通标志识别的目标图像。
S120、将所述第一图片输入至去干扰自编码器进行预处理，以过滤所述第一图片中的干扰，输出第二图片，其中，所述去干扰自编码器通过至少两类干扰样本集进行训练而得，不同类型干扰样本集中所加入的扰动方式包括下述至少两种：噪音、仿射变化、滤波模糊化、亮度变化和单色化。
在第一图片中，可能包含有交通标志、信号灯或车道线等具有交通指示作用的信息，也可能不包含。其中，包含有交通指示信息在内的第一图片通常对车辆的控制起到决定性作用。在一些情况中，交通指示牌、信号灯或车道线等交通标示上会被贴广告、标签或是叠加图贴等方式进行干扰使交通标志识别模型不能正确的识别出交通标志，从而导致违反交通规则，甚至是危害乘客人身安全以及公共交通安全。
因此，需要在将含有交通标示的图片输入到交通标志识别模型之前进行图片的预处理，将图片中可能存在的干扰信息滤除，相当于提炼出了图片中的关键物体信息。
例如，可将第一图片输入到去干扰自编码器进行预处理，即可在含有交通标志信息的第一图像中有干扰信息时将干扰信息滤除，得到第二图片，即无干扰的图像。对于不包含交通标志信息及包含交通标志信息但未添加干扰信息的第一图片，去干扰自编码器的预处理是对该图片没有较大影响的，可得到接近于原图的输出图像。其中，去干扰自编码器通过至少两类干扰样本集进行训练而获得的，不仅能过滤掉单一图片干扰处理的干扰，还可以滤掉多种干扰处理方法组合的干扰，提高对抗样本图片中的扰动过滤效果。
在每一类抗干扰样本集中，包含有至少一个样本对，每一个样本对都包含有一个原始图片，以及一个与该原始图片相对应的对抗样本。在一类抗干扰样本集中，每个对抗样本相对于相应的原始图片均进行了相同类型的扰动处理。所谓相同类型，是指采用的扰动方式的组合相同。扰动方式的组合可以包括单一一种扰动方式，或者也可以包括两种或多种扰动方式的组合。在一类抗干扰样本集中，采用的扰动方式组合相同，但每扰动方式所采用的具体参数可以相同或不同。本申请实施例中所采用的扰动方式可能由多种，可选的，扰动方式包括噪音、仿射变化、滤波模糊化、亮度变化和单色化中的至少两种。
在一种优选的实施方式中，在将第一图片输入至去干扰自编码器进行预处理之前，还可以对第一图片从颜色维度进行压缩处理，即RGB颜色信息、灰度等级或RGB颜色信息和灰度等级方面的压缩处理。这是因为交通标志的识别，主要依赖于交通标志图案的结构、形状、主体颜色，而对细节颜色并不敏感。通常情况下，交通标志在阳光下、黑暗中所采集呈现的颜色也是有差别的，所以细微颜色的差异被压缩并不影响交通标志图案的识别。图像在颜色维度进行压缩之后可以减少在图像处理过程中的运算的数据量。
S130、将所述第二图片输入至交通标志识别模型来进行识别处理。
其中，交通标志识别模型通常为基于深度学习的网络模型。
交通标志识别模型可以识别第二图像中特征信息，并判断特征信息是否属于任一个交通标志，如限速指示牌或交通灯等，以供车辆驾驶控制系统的决策模块根据交通标志识别 模型的识别结果做出控制决策，进行车辆行驶过程中的控制。
本实施例的技术方案，通过将车辆采集的视频流中的图片输入至去干扰自编码器，经过去干扰自编码器的预处理得到滤除干扰的图片，进而将无干扰的图片输入至交通标志识别模型进行识别处理，从而使得后续能够产生正确的车辆控制指令，解决了对抗样本对交通标志识别模型进行攻击导致交通标志识别错误的问题；可以降低交通图像中对抗样本的干扰，提高图像的识别正确率，提升无人驾驶或智能驾驶的安全性。
本申请实施例的技术方案，可同时适用于一些不法用户在不确定交通标志识别所使用的深度学习模型时所发起的黑盒攻击,以及在确定深度学习模型时的白盒攻击两种情况。黑盒攻击不同于白盒攻击。白盒攻击往往是在已知深度学习模型的模型结构和具体参数时有针对性的采用快速梯度信号方法(Fast Gradient Sign Method，FGSM)、CW(Clarke and Wright)、基于雅可比的显着映射方法(Jacobian-based saliency map approach，JSMA)等对抗样本算法进行白盒攻击。而黑盒攻击在不确定深度学习模型时，会通过诸如噪音、仿射变化、滤波模糊化、亮度变化和单色化等扰动方式，发起复杂多变的黑盒攻击。本申请实施例有效解决了黑盒攻击和白盒攻击的情况过滤掉了每种扰动，让交通标志识别的深度学习模型能有效进行识别过滤。
实施例二
图2a为本申请实施例二提供的一种交通图像识别方法的流程图。本实施例以上述实施例中每个可选方案为基础，提供了去干扰自编码器的训练过程。如图2a所示，本申请实施例中提供的交通图像识别方法包括如下步骤：
S210、在原始图片的基础上增加至少两种干扰，以形成至少两类干扰样本集。
其中，原始图片是未增加过干扰的图片，图片的内容为真实的交通指示灯、交通指示牌、车道线及路牌等内容。获取原始图片的途径可以是通过带有摄像功能的终端拍摄而得，也可以是在某一个视频中进行截取。在获取原始图片之后，则开始生成样本集。首先，通过加噪音、增加仿射变化、叠加滤波模糊化变化、叠加亮度变化、叠加单色化变化中的一种或多种扰动方式，对原始图片进行处理，以形成干扰图片。然后，将原始图片与干扰图片作为一个样本对，选择至少两类样本对集合作为所述干扰样本集。对于每类干扰样本集确定采用相同的扰动方式组合。
示例性的，在第一原始图片增加仿射变化和滤波模糊化变化，生成一个第一干扰图片，该第一原始图片与该第一干扰图片即为一个样本对。同样的，在其他的原始图片增加仿射变化和滤波模糊化变化生成相应的干扰图片，得到多个样本对，那么经过相同变化得到的样本对同属于一类样本对集合，即第一类样本对集合。如果，在第一原始图片中，叠加滤波模糊化变化、叠加亮度变化及叠加单色化变化，也会生成相应的干扰图片，组成相应的样本对，此时得到的样本对集合为不同于第一类样本对集合的第二类样本对集合。同理，选择在原始图片上叠加不同的种类及数量的干扰信息之后，可得到更多不同类别的样本对集合。从而， 选择至少两类样本对集合作为所述干扰样本集，以使训练样本更加全面，能够覆盖更多的扰动方式，从而能够提高对抗样本的滤除出率。
在另一种实施方式中，在通过加噪音、增加仿射变化、叠加滤波模糊化变化、叠加亮度变化、叠加单色化变化中的一种或多种扰动方式，对所述原始图片进行处理之前，还可以调整任一类型扰动方式中的至少一项扰动参数值，以形成至少两种扰动，从而增加对于同一个原始图片生成干扰图片的数量，进而增加样本对集合的数量。示例性的，调整任一类型扰动方式中的至少一项扰动参数值，以形成至少两种扰动可以包括下述至少一种：
调整放射变化中的缩放比例参数，以形成不同缩放比例的扰动；调整滤波模糊化中的模糊控制器的输入参数，以形成模糊程度不同的扰动；调整亮度变化中的亮度值，以形成不同亮度的扰动；调整单色化变化中的像素点的像素值，以形成不同颜色的扰动。当其中一种扰动方式包含多个扰动参数时，可同时改变多个参数值，形成不同的干扰图片，如同时改变放射变化中的翻转角度参数和剪切角度参数，以及亮度变化中的亮度值。
S220、将每个所述干扰样本集中的样本对分别作为输入图片和输出图片，输入至自编码器以进行训练。
自动编码器(Auto Encoders)是深度学习中常见的一种模型，其结构是一个三层神经网络结构，该结构中包括一个输入层、一个隐藏层和一个输出层，其中，输出层和输入层具有相同的维数，具体可参考图2b。具体的，输入层和输出层分别代表代表神经网络的输入层和输出层，隐藏层承担的编码器和解码器的工作，编码的过程就是从高维度的输入层转化到低维度的隐藏层的过程，反之，解码过程就是低维度的隐藏层到高维度的输出层的转化过程，因此，自编码器是个有损转化的过程，通过对比输入和输出的差别来定义损失函数。训练的过程不需要对数据进行标记，整个过程就是不断求解损失函数最小化的过程。
在本实施例中，将任一样本对中的叠加了噪声的干扰图片输入至输入层，然后，在输出层得到经过自编码器的隐藏层还原的图片，然后，将原始图片和经过还原的图片同时输入到损失函数中，根据损失函数的输出结果判断是否需要对自动编码器进行优化，当损失函数的输出结果满足预设条件时，即可停止训练过程，最终得到去干扰自编码器。
在另一种实施方式中，由于车辆采集的视频流中的图像信息是在时间上连续的有关联关系的图像信息，干扰自编码器可以为LSTM(Long Short-Term Memory，长短期记忆网络)的卷积神经网络模型。那么，干扰样本集的样本包括连续的至少两帧图片。即原始图片为至少两张连续帧图片组成的原始样本组，与原始样本组相对应的干扰图片组为在原始样本组的基础上以相同的扰动方式叠加了干扰信息的图片。其中，相同的扰动方式是指采用的扰动方式的组合相同。扰动方式的组合可以包括单一一种扰动方式，或者也可以包括两种或多种扰动方式的组合。在一类抗干扰样本集中，采用的扰动方式组合相同，但每种扰动方式所采用的具体参数可以相同或不同。本申请实施例中所采用的扰动方式可能由多种，可选的，扰动方式包括噪音、仿射变化、滤波模糊化、亮度变化和单色化中的至少两种。
在一种优选的实施方式中，在进行自编码器的训练之前，还可以对样本集中的样本图像 从颜色维度进行压缩处理，即RGB颜色信息、灰度等级或RGB颜色信息和灰度等级方面的压缩处理。这是因为交通标志的识别，主要依赖于物体的结构、形状、主体颜色，而对细节颜色并不敏感。图像在颜色维度进行压缩之后可以减少在图像处理过程中的运算的数据量。
S230、获取车辆采集的视频流并提取出所述视频流中的每帧图像作为第一图片。
S240、将所述第一图片输入至去干扰自编码器进行预处理，以过滤所述第一图片中的干扰，输出第二图片。
S250、将所述第二图片输入至交通标志识别模型来进行识别处理。
S230-S250的具体内容，可参考实施例一中的相关描述。
本实施例的技术方案，通过在原始图片通过不同的扰动方式添加干扰噪声形成不同类的干扰样本集，训练自编码器，得到了可滤掉多种干扰的去干扰自编码器，然后使用该去干扰自编码器对车辆采集到的视频流中图片进行去干扰预处理，得到过滤掉干扰的图片，将经过预处理的图片输入至交通标志识别模型进行识别处理，从而产生正确的车辆控制指令，解决了对抗样本对交通标志识别模型进行攻击导致交通标志识别错误的问题；可以降低交通图像中对抗样本的干扰，提高图像的识别正确率，提升无人驾驶或智能驾驶的安全性。
实施例三
图3示出了本申请实施例三提供的一种交通图像识别装置的结构示意图，本申请实施例可适用于抵御基于对抗样本对无人车或智能驾驶控制系统的路牌和交通指示灯识别模型的攻击的情况。
如图3所示，本申请实施例中交通图像识别装置，包括：图片采集模块310、图片预处理模块320和图片识别模块330。
其中，图片采集模块310，设置为获取车辆采集的视频流并提取出所述视频流中的每帧图像作为第一图片；图片预处理模块320，设置为将所述第一图片输入至去干扰自编码器进行预处理，以过滤所述第一图片中的干扰，输出第二图片，其中，所述去干扰自编码器通过至少两类干扰样本集进行训练而得，不同类型干扰样本集中所加入的干扰类型包括下述至少两种：噪音、仿射变化、滤波模糊化、亮度变化和单色化；图片识别模块330，设置为将所述第二图片输入至交通标志识别模型来进行识别处理。
本实施例的技术方案，通过将车辆采集的视频流中的图片输入至去干扰自编码器，经过去干扰自编码器的预处理得到滤除干扰的图片，进而将无干扰的图片输入至交通标志识别模型进行识别处理从而产生正确的车辆控制指令，解决了对抗样本对交通标志识别模型进行攻击导致交通标志识别错误的问题；可以降低交通图像中对抗样本的干扰，提高图像的识别正确率，提升无人驾驶或智能驾驶的安全性。
在一实施例中，交通图像识别装置还包括：样本集生成模块，设置为在原始图片的基础上增加至少两种干扰，以形成至少两类干扰样本集；模型训练模块，设置为将每个所述干扰样本集中的样本对分别作为输入图片和输出图片，输入至自编码器以进行训练。
在一实施例中，样本集生成模块设置为：获取原始图片；通过加噪音、增加仿射变化、叠加滤波模糊化变化、叠加亮度变化、叠加单色化变化中的一种或多种扰动方式，对所述原始图片进行处理，以形成干扰图片；将原始图片与干扰图片作为一个样本对，选择至少两类样本对集合作为所述干扰样本集。
在一实施例中，样本集生成模块，还设置为：调整任一类型扰动方式中的至少一项扰动参数值，以形成至少两种扰动。
在一实施例中，调整任一类型扰动方式中的至少一项扰动参数值，以形成至少两种扰动包括下述至少一种：调整放射变化中的缩放比例参数，以形成不同缩放比例的扰动；调整滤波模糊化中的模糊控制器的输入参数，以形成模糊程度不同的扰动；调整亮度变化中的亮度值，以形成不同亮度的扰动；调整单色化变化中的像素点的像素值，以形成不同颜色的扰动。
在一实施例中，自编码器的输入层和输出层结构相同，以使输出图片与原始图片具有相同的分辨率。
在一实施例中，交通图像识别装置还包括图像压缩模块，设置为在将所述第一图片输入至去干扰自编码器进行预处理之前，对所述第一图片从颜色维度进行压缩处理。
在一实施例中，所述去干扰自编码器为LSTM的卷积神经网络模型，所述干扰样本集包括连续的至少两帧图片。
本申请实施例所提供的交通图像识别装置可执行本申请任意实施例所提供的交通图像识别方法，具备执行方法相应的功能模块和有益效果。
实施例四
图4是本申请实施例四中的计算机设备的结构示意图。图4示出了适于用来实现本申请实施方式的示例性计算机设备412的框图。图4显示的计算机设备412仅仅是一个示例，不应对本申请实施例的功能和使用范围带来任何限制。
如图4所示，计算机设备412以通用计算设备的形式表现。计算机设备412的组件可以包括但不限于：一个或者多个处理器或者处理单元416，系统存储器428，连接不同系统组件(包括系统存储器428和处理单元416)的总线418。
总线418表示几类总线结构中的一种或多种，包括存储器总线或者存储器控制器，外围总线，图形加速端口，处理器或者使用多种总线结构中的任意总线结构的局域总线。举例来说，这些体系结构包括但不限于工业标准体系结构(ISA)总线，微通道体系结构(MAC)总线，增强型ISA总线、视频电子标准协会(VESA)局域总线以及外围组件互连(PCI)总线。
计算机设备412典型地包括多种计算机系统可读介质。这些介质可以是任何能够被计算机设备412访问的可用介质，包括易失性和非易失性介质，可移动的和不可移动的介质。
系统存储器428可以包括易失性存储器形式的计算机系统可读介质，例如随机存取存储器(RAM)430和/或高速缓存存储器432。计算机设备412可以进一步包括其它可移动/不可 移动的、易失性/非易失性计算机系统存储介质。仅作为举例，存储系统434可以用于读写不可移动的、非易失性磁介质(图4未显示，通常称为“硬盘驱动器”)。尽管图4中未示出，可以提供用于对可移动非易失性磁盘(例如“软盘”)读写的磁盘驱动器，以及对可移动非易失性光盘(例如CD-ROM,DVD-ROM或者其它光介质)读写的光盘驱动器。在这些情况下，每个驱动器可以通过一个或者多个数据介质接口与总线418相连。存储器428可以包括至少一个程序产品，该程序产品具有一组(例如至少一个)程序模块，这些程序模块被配置以执行本申请每个实施例的功能。
具有一组(至少一个)程序模块442的程序/实用工具440，可以存储在例如存储器428中，这样的程序模块442包括但不限于操作系统、一个或者多个应用程序、其它程序模块以及程序数据，这些示例中的每一个或某种组合中可能包括网络环境的实现。程序模块442通常执行本申请所描述的实施例中的功能和/或方法。
计算机设备412也可以与一个或多个外部设备414(例如键盘、指向设备、显示器424等)通信，还可与一个或者多个使得用户能与该计算机设备412交互的设备通信，和/或与使得该计算机设备412能与一个或多个其它计算设备进行通信的任何设备(例如网卡，调制解调器等等)通信。这种通信可以通过输入/输出(I/O)接口422进行。并且，计算机设备412还可以通过网络适配器420与一个或者多个网络(例如局域网(LAN)，广域网(WAN)和/或公共网络，例如因特网)通信。如图所示，网络适配器420通过总线418与计算机设备412的其它模块通信。应当明白，尽管图4中未示出，可以结合计算机设备412使用其它硬件和/或软件模块，包括但不限于：微代码、设备驱动器、冗余处理单元、外部磁盘驱动阵列、RAID系统、磁带驱动器以及数据备份存储系统等。
处理单元416通过运行存储在系统存储器428中的程序，从而执行每种功能应用以及数据处理，例如实现本申请实施例所提供的交通图像识别方法，该方法主要包括：获取车辆采集的视频流并提取出所述视频流中的每帧图像作为第一图片；将所述第一图片输入至去干扰自编码器进行预处理，以过滤所述第一图片中的干扰，输出第二图片，其中，所述去干扰自编码器通过至少两类干扰样本集进行训练而得，不同类型干扰样本集中所加入的扰动方式包括下述至少两种：噪音、仿射变化、滤波模糊化、亮度变化和单色化；将所述第二图片输入至交通标志识别模型来进行识别处理。
实施例五
本申请实施例五还提供了一种计算机可读存储介质，其上存储有计算机程序，该程序被处理器执行时实现如本申请实施例所提供的交通图像识别方法，该方法主要包括：获取车辆采集的视频流并提取出所述视频流中的每帧图像作为第一图片；将所述第一图片输入至去干扰自编码器进行预处理，以过滤所述第一图片中的干扰，输出第二图片，其中，所述去干扰自编码器通过至少两类干扰样本集进行训练而得，不同类型干扰样本集中所加入的扰动方式包括下述至少两种：噪音、仿射变化、滤波模糊化、亮度变化和单色化；将所述第二图片输 入至交通标志识别模型来进行识别处理。
本申请实施例的计算机存储介质，可以采用一个或多个计算机可读的介质的任意组合。计算机可读介质可以是计算机可读信号介质或者计算机可读存储介质。计算机可读存储介质例如可以是——但不限于——电、磁、光、电磁、红外线、或半导体的系统、装置或器件，或者任意以上的组合。计算机可读存储介质的更具体的例子(非穷举的列表)包括：具有一个或多个导线的电连接、便携式计算机磁盘、硬盘、随机存取存储器(RAM)、只读存储器(ROM)、可擦式可编程只读存储器(EPROM或闪存)、光纤、便携式紧凑磁盘只读存储器(CD-ROM)、光存储器件、磁存储器件、或者上述的任意合适的组合。在本文件中，计算机可读存储介质可以是任何包含或存储程序的有形介质，该程序可以被指令执行系统、装置或者器件使用或者与其结合使用。
计算机可读的信号介质可以包括在基带中或者作为载波一部分传播的数据信号，其中承载了计算机可读的程序代码。这种传播的数据信号可以采用多种形式，包括但不限于电磁信号、光信号或上述的任意合适的组合。计算机可读的信号介质还可以是计算机可读存储介质以外的任何计算机可读介质，该计算机可读介质可以发送、传播或者传输用于由指令执行系统、装置或者器件使用或者与其结合使用的程序。
计算机可读介质上包含的程序代码可以用任何适当的介质传输，包括——但不限于无线、电线、光缆、RF等等，或者上述的任意合适的组合。
可以以一种或多种程序设计语言或其组合来编写用于执行本申请操作的计算机程序代码，所述程序设计语言包括面向对象的程序设计语言—诸如Java、Smalltalk、C++，还包括常规的过程式程序设计语言—诸如”C”语言或类似的程序设计语言。程序代码可以完全地在用户计算机上执行、部分地在用户计算机上执行、作为一个独立的软件包执行、部分在用户计算机上部分在远程计算机上执行、或者完全在远程计算机或服务器上执行。在涉及远程计算机的情形中，远程计算机可以通过任意种类的网络——包括局域网(LAN)或广域网(WAN)—连接到用户计算机，或者，可以连接到外部计算机(例如利用因特网服务提供商来通过因特网连接)。

Claims (12)

  1. 一种交通图像识别方法，包括：
    获取车辆采集的视频流并提取出所述视频流中的每帧图像作为第一图片；
    将所述第一图片输入至去干扰自编码器进行预处理，以过滤所述第一图片中的干扰，输出第二图片，其中，所述去干扰自编码器通过至少两类干扰样本集进行训练而得，不同类型干扰样本集中所加入的扰动方式包括下述至少两种：噪音、仿射变化、滤波模糊化、亮度变化和单色化；
    将所述第二图片输入至交通标志识别模型来进行识别处理。
  2. 根据权利要求1所述的方法，还包括：
    在原始图片的基础上增加至少两种干扰，以形成至少两类干扰样本集；
    将每个所述干扰样本集中的样本对分别作为输入图片和输出图片，输入至自编码器以进行训练。
  3. 根据权利要求2所述的方法，其中，在原始图片的基础上增加至少两种干扰，以形成至少两类干扰样本集包括：
    获取原始图片；
    通过加噪音、增加仿射变化、叠加滤波模糊化变化、叠加亮度变化、叠加单色化变化中的至少一种扰动方式，对所述原始图片进行处理，以形成干扰图片；
    将原始图片与干扰图片作为一个样本对，选择至少两类样本对集合作为所述干扰样本集。
  4. 根据权利要求3所述的方法，通过加噪音、增加仿射变化、叠加滤波模糊化变化、叠加亮度变化、叠加单色化变化中的至少一种扰动方式，对所述原始图片进行处理之前，还包括：
    调整任一类型扰动方式中的至少一项扰动参数值，以形成至少两种扰动。
  5. 根据权利要求4所述的方法，其中，调整任一类型扰动方式中的至少一项扰动参数值，以形成至少两种扰动包括下述至少一种：
    调整放射变化中的缩放比例参数，以形成不同缩放比例的扰动；
    调整滤波模糊化中的模糊控制器的输入参数，以形成模糊程度不同的扰动；
    调整亮度变化中的亮度值，以形成不同亮度的扰动；
    调整单色化变化中的像素点的像素值，以形成不同颜色的扰动。
  6. 根据权利要求2-5中任一项所述的方法，其中，所述自编码器的输入层和输出层结构相同，以使输出图片与原始图片具有相同的分辨率。
  7. 根据权利要求6所述的方法，在将所述第一图片输入至去干扰自编码器进行预处理之前，还包括：
    对所述第一图片从颜色维度进行压缩处理。
  8. 根据权利要求1所述的方法，其中，所述去干扰自编码器为LSTM的卷积神经网络模型，所述干扰样本集包括连续的至少两帧图片。
  9. 一种交通图像识别装置，包括：
    图片采集模块，设置为获取车辆采集的视频流并提取出所述视频流中的每帧图像作为第一图片；
    图片预处理模块，设置为将所述第一图片输入至去干扰自编码器进行预处理，以过滤所述第一图片中的干扰，输出第二图片，其中，所述去干扰自编码器通过至少两类干扰样本集进行训练而得，不同类型干扰样本集中所加入的干扰类型包括下述至少两种：噪音、仿射变化、滤波模糊化、亮度变化和单色化；
    图片识别模块，设置为将所述第二图片输入至交通标志识别模型来进行识别处理。
  10. 根据权利要求9所述的装置，还包括：
    样本集生成模块，设置为在原始图片的基础上增加至少两种干扰，以形成至少两类干扰样本集；
    模型训练模块，设置为将每个所述干扰样本集中的样本对分别作为输入图片和输出图片，输入至自编码器以进行训练。
  11. 一种计算机设备，包括：
    至少一个处理器；
    存储装置，设置为存储至少一个程序；
    当所述至少一个程序被所述至少一个处理器执行，使得所述至少一个处理器实现如权利要求1-8中任一项所述的交通图像识别方法。
  12. 一种计算机可读存储介质，其上存储有计算机程序，所述计算机程序被处理器执行时实现如权利要求1-8中任一项所述的交通图像识别方法。
PCT/CN2019/102027 2019-02-25 2019-08-22 交通图像识别方法、装置、计算机设备和介质 WO2020173056A1 (zh)

Priority Applications (4)

Application Number Priority Date Filing Date Title
EP19916553.1A EP3786835A4 (en) 2019-02-25 2019-08-22 METHOD AND APPARATUS FOR TRAFFIC IMAGE RECOGNITION, AND COMPUTER DEVICE AND MEDIA
JP2020568528A JP2022521448A (ja) 2019-02-25 2019-08-22 交通画像認識方法、装置、コンピュータデバイスおよび媒体
KR1020207035694A KR20210031427A (ko) 2019-02-25 2019-08-22 교통 이미지를 인식하는 방법, 장치, 컴퓨터 기기 및 매체
US17/114,076 US20210117705A1 (en) 2019-02-25 2020-12-07 Traffic image recognition method and apparatus, and computer device and medium

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
CN201910138054.7 2019-02-25
CN201910138054.7A CN109886210B (zh) 2019-02-25 2019-02-25 一种交通图像识别方法、装置、计算机设备和介质

Related Child Applications (1)

Application Number Title Priority Date Filing Date
US17/114,076 Continuation US20210117705A1 (en) 2019-02-25 2020-12-07 Traffic image recognition method and apparatus, and computer device and medium

Publications (1)

Publication Number Publication Date
WO2020173056A1 true WO2020173056A1 (zh) 2020-09-03

Family

ID=66929338

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2019/102027 WO2020173056A1 (zh) 2019-02-25 2019-08-22 交通图像识别方法、装置、计算机设备和介质

Country Status (6)

Country Link
US (1) US20210117705A1 (zh)
EP (1) EP3786835A4 (zh)
JP (1) JP2022521448A (zh)
KR (1) KR20210031427A (zh)
CN (1) CN109886210B (zh)
WO (1) WO2020173056A1 (zh)

Cited By (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN112241532A (zh) * 2020-09-17 2021-01-19 北京科技大学 一种基于雅可比矩阵生成与检测恶性对抗样本的方法
CN112990015A (zh) * 2021-03-16 2021-06-18 北京智源人工智能研究院 一种病变细胞自动识别方法、装置和电子设备
CN114004757A (zh) * 2021-10-14 2022-02-01 大族激光科技产业集团股份有限公司 去除工业图像中干扰的方法、系统、设备和存储介质

Families Citing this family (16)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN109886210B (zh) * 2019-02-25 2022-07-19 百度在线网络技术（北京）有限公司 一种交通图像识别方法、装置、计算机设备和介质
CN110717028B (zh) * 2019-10-18 2022-02-15 支付宝(杭州)信息技术有限公司 一种剔除干扰问题对的方法及系统
CN112906424B (zh) * 2019-11-19 2023-10-31 上海高德威智能交通系统有限公司 图像识别方法、装置及设备
KR102382324B1 (ko) * 2019-11-21 2022-04-01 연세대학교 산학협력단 신경 파형 분류 장치 및 방법
CN111191717B (zh) * 2019-12-30 2022-05-10 电子科技大学 一种基于隐空间聚类的黑盒对抗样本生成算法
CN111553952A (zh) * 2020-05-08 2020-08-18 中国科学院自动化研究所 基于生存对抗的工业机器人视觉图像识别方法及系统
CN111783604A (zh) * 2020-06-24 2020-10-16 中国第一汽车股份有限公司 基于目标识别的车辆控制方法、装置、设备及车辆
CN111899199B (zh) * 2020-08-07 2024-03-19 深圳市捷顺科技实业股份有限公司 一种图像处理方法、装置、设备及存储介质
CN111967368B (zh) * 2020-08-12 2022-03-11 广州小鹏自动驾驶科技有限公司 一种交通灯识别的方法和装置
JP6968475B1 (ja) * 2021-06-03 2021-11-17 望 窪田 情報処理方法、プログラム及び情報処理装置
CN113537463A (zh) * 2021-07-02 2021-10-22 北京航空航天大学 基于数据扰动的对抗样本防御方法与装置
CN113255609B (zh) * 2021-07-02 2021-10-29 智道网联科技(北京)有限公司 基于神经网络模型的交通标识识别方法及装置
EP4120136A1 (en) * 2021-07-14 2023-01-18 Volkswagen Aktiengesellschaft Method for automatically executing a vehicle function, method for training a machine learning defense model and defense unit for a vehicle
CN113537494B (zh) * 2021-07-23 2022-11-11 江南大学 一种基于黑盒场景的图像对抗样本生成方法
US11887475B1 (en) * 2021-12-13 2024-01-30 Ford Global Technologies, Llc Systems and methods for controlling a programmable traffic light
CN115588131B (zh) * 2022-09-30 2024-02-06 北京瑞莱智慧科技有限公司 模型鲁棒性检测方法、相关装置及存储介质

Citations (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20130077830A1 (en) * 2011-09-23 2013-03-28 Ricoh Company, Ltd. Traffic sign detecting method and traffic sign detecting device
CN106022268A (zh) * 2016-05-23 2016-10-12 广州鹰瞰信息科技有限公司 一种限速标识的识别方法和装置
CN106919939A (zh) * 2017-03-14 2017-07-04 潍坊学院 一种交通标识牌跟踪识别方法及系统
CN107122737A (zh) * 2017-04-26 2017-09-01 聊城大学 一种道路交通标志自动检测识别方法
CN107679508A (zh) * 2017-10-17 2018-02-09 广州汽车集团股份有限公司 交通标志检测识别方法、装置及系统
CN109886210A (zh) * 2019-02-25 2019-06-14 百度在线网络技术（北京）有限公司 一种交通图像识别方法、装置、计算机设备和介质

Family Cites Families (18)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
JPH05128250A (ja) * 1991-11-08 1993-05-25 Toshiba Corp 画像認識装置
JP2004354251A (ja) * 2003-05-29 2004-12-16 Nidek Co Ltd 欠陥検査装置
JP5082512B2 (ja) * 2007-03-08 2012-11-28 富士ゼロックス株式会社 情報処理装置、画像処理装置、画像符号化装置、情報処理プログラム、画像処理プログラム及び画像符号化プログラム
CN105590088A (zh) * 2015-09-17 2016-05-18 重庆大学 一种基于稀疏自编码和稀疏表示进行交通标志识别的方法
CN105139342A (zh) * 2015-09-29 2015-12-09 天脉聚源（北京）教育科技有限公司 一种图片缩放的方法和装置
AU2017210160A1 (en) * 2016-01-18 2018-08-02 Waveshift Llc Evaluating and reducing myopiagenic effects of electronic displays
JP6688090B2 (ja) * 2016-01-22 2020-04-28 株式会社デンソーテン 物体認識装置および物体認識方法
CN106127702B (zh) * 2016-06-17 2018-08-14 兰州理工大学 一种基于深度学习的图像去雾方法
CN106529589A (zh) * 2016-11-03 2017-03-22 温州大学 采用降噪堆叠自动编码器网络的视觉目标检测方法
CN107571867B (zh) * 2017-09-05 2019-11-08 百度在线网络技术（北京）有限公司 用于控制无人驾驶车辆的方法和装置
CN108122209B (zh) * 2017-12-14 2020-05-15 浙江捷尚视觉科技股份有限公司 一种基于对抗生成网络的车牌去模糊方法
CN108416752B (zh) * 2018-03-12 2021-09-07 中山大学 一种基于生成式对抗网络进行图像去运动模糊的方法
CN108537133A (zh) * 2018-03-16 2018-09-14 江苏经贸职业技术学院 一种基于监督学习深度自编码器的人脸重构方法
CN108520503B (zh) * 2018-04-13 2020-12-22 湘潭大学 一种基于自编码器和生成对抗网络修复人脸缺损图像的方法
CN108710831B (zh) * 2018-04-24 2021-09-21 华南理工大学 一种基于机器视觉的小数据集人脸识别算法
CN108961217B (zh) * 2018-06-08 2022-09-16 南京大学 一种基于正例训练的表面缺陷检测方法
CN109191402B (zh) * 2018-09-03 2020-11-03 武汉大学 基于对抗生成神经网络的图像修复方法和系统
US11106925B2 (en) * 2018-10-25 2021-08-31 Intel Corporation Computer-assisted or autonomous driving traffic sign recognition method and apparatus

Patent Citations (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20130077830A1 (en) * 2011-09-23 2013-03-28 Ricoh Company, Ltd. Traffic sign detecting method and traffic sign detecting device
CN106022268A (zh) * 2016-05-23 2016-10-12 广州鹰瞰信息科技有限公司 一种限速标识的识别方法和装置
CN106919939A (zh) * 2017-03-14 2017-07-04 潍坊学院 一种交通标识牌跟踪识别方法及系统
CN107122737A (zh) * 2017-04-26 2017-09-01 聊城大学 一种道路交通标志自动检测识别方法
CN107679508A (zh) * 2017-10-17 2018-02-09 广州汽车集团股份有限公司 交通标志检测识别方法、装置及系统
CN109886210A (zh) * 2019-02-25 2019-06-14 百度在线网络技术（北京）有限公司 一种交通图像识别方法、装置、计算机设备和介质

Cited By (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN112241532A (zh) * 2020-09-17 2021-01-19 北京科技大学 一种基于雅可比矩阵生成与检测恶性对抗样本的方法
CN112241532B (zh) * 2020-09-17 2024-02-20 北京科技大学 一种基于雅可比矩阵生成与检测恶性对抗样本的方法
CN112990015A (zh) * 2021-03-16 2021-06-18 北京智源人工智能研究院 一种病变细胞自动识别方法、装置和电子设备
CN112990015B (zh) * 2021-03-16 2024-03-19 北京智源人工智能研究院 一种病变细胞自动识别方法、装置和电子设备
CN114004757A (zh) * 2021-10-14 2022-02-01 大族激光科技产业集团股份有限公司 去除工业图像中干扰的方法、系统、设备和存储介质
CN114004757B (zh) * 2021-10-14 2024-04-05 大族激光科技产业集团股份有限公司 去除工业图像中干扰的方法、系统、设备和存储介质

Also Published As

Publication number Publication date
EP3786835A1 (en) 2021-03-03
CN109886210B (zh) 2022-07-19
US20210117705A1 (en) 2021-04-22
KR20210031427A (ko) 2021-03-19
JP2022521448A (ja) 2022-04-08
CN109886210A (zh) 2019-06-14
EP3786835A4 (en) 2022-01-26

Similar Documents

Publication Publication Date Title
WO2020173056A1 (zh) 交通图像识别方法、装置、计算机设备和介质
CN110969160B (zh) 一种基于深度学习的车牌图像校正、识别方法及系统
WO2021244621A1 (zh) 基于全局引导选择性上下文网络的场景语义解析方法
CN109902617B (zh) 一种图片识别方法、装置、计算机设备和介质
CN111899515B (zh) 一种基于智慧道路边缘计算网关的车辆检测系统
CN113723377A (zh) 一种基于ld-ssd网络的交通标志检测方法
CN114817991B (zh) 一种车联网图像脱敏方法和系统
CN114596548A (zh) 目标检测方法、装置、计算机设备及计算机可读存储介质
CN117218622A (zh) 路况检测方法、电子设备及存储介质
CN116052090A (zh) 图像质量评估方法、模型训练方法、装置、设备及介质
CN112115928B (zh) 基于违停车辆标签的神经网络的训练方法和检测方法
CN112016534B (zh) 车辆违停检测的神经网络的训练方法、检测方法和装置
CN112052863B (zh) 一种图像检测方法及装置、计算机存储介质、电子设备
CN114820676A (zh) 一种设备运行状态识别方法及装置
CN111369753A (zh) 非机动车监测方法及相关产品
Sankaranarayanan et al. Improved Vehicle Detection Accuracy and Processing Time for Video Based ITS Applications
CN115578246B (zh) 基于风格迁移的非对齐可见光和红外模态融合目标检测方法
CN111476353B (zh) 一种引入显著性的gan图像超分辨率方法
CN116052031A (zh) 车辆违停检测方法、装置、存储介质及电子设备
Zhang et al. Research on real-time detection algorithm for pedestrian and vehicle in foggy weather based on lightweight XM-YOLOViT
CN118072287A (zh) 障碍物检测方法、训练方法及装置
CN117636060A (zh) 一种基于翼身融合特征的无人机检测方法
Lin et al. Vehicle and pedestrian detection method based on improved YOLOX in foggy environment
CN117409380A (zh) 识别车辆的方法和装置
CN115205805A (zh) 一种基于递进式特征学习技术的车型识别方法和系统

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: 19916553

Country of ref document: EP

Kind code of ref document: A1

ENP Entry into the national phase

Ref document number: 2019916553

Country of ref document: EP

Effective date: 20201125

ENP Entry into the national phase

Ref document number: 2020568528

Country of ref document: JP

Kind code of ref document: A

NENP Non-entry into the national phase

Ref country code: DE