WO2020108023A1 - 视频动作分类的方法、装置、计算机设备和存储介质 - Google Patents

视频动作分类的方法、装置、计算机设备和存储介质 Download PDF

Info

Publication number
WO2020108023A1
WO2020108023A1 PCT/CN2019/106250 CN2019106250W WO2020108023A1 WO 2020108023 A1 WO2020108023 A1 WO 2020108023A1 CN 2019106250 W CN2019106250 W CN 2019106250W WO 2020108023 A1 WO2020108023 A1 WO 2020108023A1
Authority
WO
WIPO (PCT)
Prior art keywords
optical flow
video frames
video
information corresponding
group
Prior art date
Legal status (The legal status 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 status listed.)
Ceased
Application number
PCT/CN2019/106250
Other languages
English (en)
French (fr)
Inventor
张志伟
李岩
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Beijing Dajia Internet Information Technology Co Ltd
Original Assignee
Beijing Dajia Internet Information Technology Co Ltd
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 Beijing Dajia Internet Information Technology Co Ltd filed Critical Beijing Dajia Internet Information Technology Co Ltd
Publication of WO2020108023A1 publication Critical patent/WO2020108023A1/zh
Priority to US17/148,106 priority Critical patent/US20210133457A1/en
Anticipated expiration legal-status Critical
Ceased legal-status Critical Current

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; 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 OR CALCULATING; 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
    • G06F18/213—Feature extraction, e.g. by transforming the feature space; Summarisation; Mappings, e.g. subspace methods
    • G06F18/2134—Feature extraction, e.g. by transforming the feature space; Summarisation; Mappings, e.g. subspace methods based on separation criteria, e.g. independent component analysis
    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/24—Classification techniques
    • G06F18/243—Classification techniques relating to the number of classes
    • G06F18/2431—Multiple classes
    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; 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 OR CALCULATING; 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 OR CALCULATING; 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 OR CALCULATING; 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
    • G06N3/0464—Convolutional networks [CNN, ConvNet]
    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; 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
    • G06N3/09—Supervised learning
    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; 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
    • G06V10/80—Fusion, i.e. combining data from various sources at the sensor level, preprocessing level, feature extraction level or classification level
    • G06V10/806—Fusion, i.e. combining data from various sources at the sensor level, preprocessing level, feature extraction level or classification level of extracted features
    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; 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 OR CALCULATING; COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V20/00—Scenes; Scene-specific elements
    • G06V20/40—Scenes; Scene-specific elements in video content
    • G06V20/41—Higher-level, semantic clustering, classification or understanding of video scenes, e.g. detection, labelling or Markovian modelling of sport events or news items
    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V20/00—Scenes; Scene-specific elements
    • G06V20/40—Scenes; Scene-specific elements in video content
    • G06V20/46—Extracting features or characteristics from the video content, e.g. video fingerprints, representative shots or key frames

Definitions

  • This application relates to the technical field of machine learning models, and in particular to a method, device, computer equipment, and storage medium for classifying video actions.
  • the relevant personnel in the short video platform can view the short video and classify the actions of the objects in the short video according to subjective ideas, such as dancing, climbing trees, drinking Water, etc., and then the relevant personnel can add corresponding tags to the short video according to the classification result.
  • embodiments of the present application provide a method and device for classifying video actions.
  • a method for video action classification including:
  • the classification category information corresponding to the video to be classified is determined.
  • an apparatus for classifying video actions including:
  • the first determining unit is configured to acquire the video to be classified and determine a plurality of video frames in the video to be classified;
  • the first input unit is configured to input the multiple video frames into the optical flow replacement module in the trained optimized video action classification model to obtain optical flow feature information corresponding to the multiple video frames;
  • Multiple video frames are input into the three-dimensional convolutional neural module in the trained optimized video action classification model to obtain spatial feature information corresponding to the multiple video frames;
  • the second determining unit is configured to determine classification category information corresponding to the video to be classified based on the optical flow feature information and the spatial feature information.
  • a computer device including:
  • Memory for storing processor executable instructions
  • the processor is configured to:
  • classification category information corresponding to the video to be classified is determined.
  • a non-transitory computer-readable storage medium when instructions in the storage medium are executed by a processor of a computer device, the computer device can perform a video action Classification method, the method includes:
  • classification category information corresponding to the video to be classified is determined.
  • a computer program product which when the computer program product is executed by a processor of a computer device, enables the computer device to perform a method of video action classification, the method include:
  • classification category information corresponding to the video to be classified is determined.
  • multiple video frames of the video to be classified can be directly input into the trained optimized video action classification model, and the trained optimized video action classification model can automatically classify the classified video, and Finally, the classification category information corresponding to the video to be classified is obtained, and the efficiency of classification processing is improved.
  • the video frame is directly used as the input of the optical flow substitution module in the model.
  • the optical flow substitution module can directly extract optical flow feature information corresponding to multiple video frames of the video to be classified, and determine the classification category information corresponding to the video to be classified based on the optical flow feature information To further improve the efficiency of classification processing.
  • Fig. 1 is a flow chart of a method for classifying video actions according to an exemplary embodiment
  • Fig. 2 is a flow chart showing a method for classifying video actions according to an exemplary embodiment
  • Fig. 3 is a flowchart of a method for training an optimized video action classification model according to an exemplary embodiment
  • Fig. 4 is a flowchart of a method for training an optimized video action classification model according to an exemplary embodiment
  • Fig. 5 is a block diagram of a device for classifying video actions according to an exemplary embodiment
  • Fig. 6 is a block diagram of a device for video action classification according to an exemplary embodiment.
  • the short video platform needs to classify the actions of the objects in the short video, such as dancing, climbing trees, drinking water, etc., and then add corresponding tags to the short video according to the classification results .
  • a method for automatically classifying short videos is provided.
  • Fig. 1 is a flowchart of a method for classifying video actions according to an exemplary embodiment. As shown in Fig. 1, the method for classifying video actions is used in a server of a short video platform, and includes the following steps.
  • step S110 the video to be classified is acquired, and multiple video frames in the video to be classified are determined.
  • the server of the short video platform can receive a large number of short videos uploaded by the user, and any short video can be used as the video to be classified, so the server can obtain the video to be classified. Since a video to be classified is composed of many video frames, it is not necessary to use all the video frames in a video to be classified for subsequent steps, so the server can extract the pre Set the number of multiple video frames.
  • the server may randomly extract a preset number of multiple video frames from all video frames in a video to be classified. The preset number can be set according to the empirical value, for example, the preset number is 10, the preset number is 5, and so on.
  • step S120 multiple video frames are input to the optical flow replacement module in the trained optimized video action classification model to obtain optical flow feature information corresponding to the multiple video frames.
  • a video motion classification model may be trained and optimized in advance for classification processing of the video to be classified.
  • the optimized video action classification model includes multiple functional modules, each of which has a different role.
  • the optimized video action classification model may include an optical flow replacement module, a three-dimensional convolutional neural module, and a first classifier module.
  • the optical flow substitution module is used to extract optical flow characteristic information corresponding to multiple video frames. As shown in FIG. 2, when the server inputs multiple video frames into the optical flow substitution module in the trained optimized video action classification model, the optical flow substitution module can output optical flow feature information corresponding to the multiple video frames. Among them, the optical flow characteristic information indicates the motion vector corresponding to the object included in the multiple video frames, that is, in what direction does the object move in the first video frame of the multiple video frames to the last shot The position in the video frame.
  • step S130 multiple video frames are input into the three-dimensional convolutional neural module in the trained optimized video action classification model to obtain spatial feature information corresponding to the multiple video frames.
  • the three-dimensional convolutional neural module may include a C3D (3 Dimensions Convolution, three-dimensional convolution) module.
  • the three-dimensional convolutional neural module is used to extract spatial feature information corresponding to multiple video frames.
  • the 3D convolutional neural module can output spatial feature information corresponding to the multiple video frames.
  • the spatial feature information indicates the position of the object included in multiple video frames in each video frame
  • the spatial feature information may be composed of a set of three-dimensional information
  • the two-dimensional in the three-dimensional information may indicate the position of the object in one video frame
  • the last dimension can represent the shooting moment corresponding to the video frame.
  • step S140 based on the optical flow feature information and the spatial feature information, the classification category information corresponding to the video to be classified is determined.
  • the server may perform feature fusion on the optical flow feature information and spatial feature information.
  • the optical flow feature information and the spatial feature information can be feature-fused through the CONCAT statement, and the fused optical flow feature information and the spatial feature information can be input into the first classifier module, and then the first classifier module
  • the classification category information corresponding to the optical flow characteristic information and the spatial characteristic information is output as the classification category information corresponding to the video to be classified to realize end-to-end classification processing.
  • the method provided in this embodiment of the present application may further include:
  • step S310 the video action classification model is trained based on the training sample, where the training sample includes multiple sets of video frames and standard classification category information corresponding to each set of video frames, and the video action classification model includes a three-dimensional convolutional neural module and optical flow Module.
  • step S320 multiple groups of video frames are respectively input to the trained optical flow module to determine reference optical flow feature information corresponding to each group of video frames.
  • step S330 an optimized video motion classification model is established based on the trained three-dimensional convolutional neural module, the preset optical flow substitution module, and the first classifier module.
  • step S340 the optimized video action classification model is trained based on multiple sets of video frames, the standard classification category information corresponding to each set of video frames and the reference optical flow feature information, and the trained optimized video action classification model is obtained.
  • the optimized video action classification model before using the trained optimized video action classification model to classify the classified video, the optimized video action classification model needs to be trained in advance.
  • the process of training the optimized video action classification model can be divided into two stages. In the first stage, the video action classification model can be trained based on the training samples. In the second stage, multiple groups of video frames can be input to the trained optical flow module to determine the reference optical flow feature information corresponding to each group of video frames. Based on the trained 3D convolutional neural module and the preset optical flow, The module and the first classifier module establish an optimized video action classification model, and train the optimized video action classification model based on multiple sets of video frames, the standard classification category information corresponding to each set of video frames, and reference optical flow feature information. Optimized video action classification model.
  • a video motion classification model can be established based on the three-dimensional convolutional neural module, optical flow module, and second classifier module.
  • the three-dimensional convolutional neural module is used to extract spatial feature information corresponding to a set of video frames
  • the optical flow module is used to extract optical flow feature information corresponding to a set of video frames
  • the second classifier module is used to correspond to a set of video frames. Spatial feature information and optical flow feature information to determine the prediction classification category information corresponding to a group of video frames.
  • multiple sets of video frames in the training samples are input into the three-dimensional convolutional neural module.
  • the three-dimensional convolutional neural module can extract the spatial feature information corresponding to each group of video frames, and can also pass the video action classification model.
  • the optical flow map corresponding to each group of video frames is determined separately, and the optical flow map corresponding to each group of video frames is input into the optical flow module, and the optical flow module can output optical flow characteristic information corresponding to each group of video frames .
  • the spatial feature information and optical flow feature information corresponding to each group of video frames can be feature-fused, and the spatial feature information and optical flow feature information corresponding to each group of video frames after fusion can be input into the second classifier module, and the second classifier The module can output the prediction classification category information corresponding to each group of video frames.
  • the standard classification category information corresponding to each group of video frames in the training sample is used as supervision information to determine the difference information between the predicted classification category information and the standard classification category information corresponding to each group of video frames. Then, the weight parameters in the video action classification model can be adjusted based on the difference information corresponding to each group of video frames. Subsequently, the above process may be repeatedly performed until it is determined that the video action classification model converges, and the trained video action classification model is obtained.
  • the difference information may be a cross entropy distance. The calculation formula of the cross entropy distance can be seen in formula 1.
  • loss entropy is the cross entropy distance
  • y is standard category information.
  • the optical flow module in the video action classification model has also been trained, that is, the optical flow module after training can be accurately extracted Optical flow characteristic information corresponding to each group of video frames. Therefore, the reference optical flow feature information output by the converged optical flow module can be added as training information to the training samples for subsequent training of other modules.
  • the weight parameter in the optical flow module may be frozen, and the adjustment of the weight parameter in the optical flow module is no longer continued.
  • the three-dimensional convolutional neural module, the preset optical flow replacement module, and the first classifier module can be used as modules in the optimized video action classification model to train the optimized video action classification model.
  • the training of the three-dimensional convolutional neural module can be continued, so that the accuracy of the results output by the three-dimensional convolutional neural module is getting higher and higher.
  • the optical flow replacement module can also be trained, so that the light The stream replacement module can replace the optical flow module to extract the optical flow characteristic information corresponding to each group of video frames.
  • the optimized video action classification model may be trained based on multiple sets of video frames, standard classification category information corresponding to each set of video frames and reference optical flow feature information to obtain the optimized video action classification after training model.
  • step S340 may include: inputting multiple groups of video frames to the optical flow replacement module to obtain predicted optical flow feature information corresponding to each group of video frames; based on the reference optical flow corresponding to each group of video frames Feature information and predicted optical flow feature information to determine the optical flow loss information corresponding to each group of video frames; input multiple groups of video frames to the trained 3D convolutional neural module to obtain reference space feature information corresponding to each group of video frames; Input the predicted optical flow feature information and reference spatial feature information corresponding to each group of video frames into the first classifier module to determine the predicted classification category information corresponding to each group of video frames; based on the standard classification category information corresponding to each group of video frames and Predict the classification category information and determine the classification loss information corresponding to each group of video frames; based on the optical flow loss information and classification loss information corresponding to each group of video frames, adjust the weighting parameters in the optical flow substitution module, based on the correspondence of each group of video frames Adjust the weighting parameter in the first classifier module.
  • multiple sets of video frames can be directly input into the optical flow substitution module, without the need to optimize the video action classification model in advance, based on the multiple sets of video frames alone, the optical flow map corresponding to each set of video frames is determined separately.
  • the optical flow replacement module can directly take multiple sets of video frames as input, without the need to take the optical flow map as input.
  • the optical flow substitution module may output the predicted optical flow characteristic information corresponding to each group of video frames.
  • the reference optical flow characteristic information corresponding to each group of video frames Since the reference optical flow characteristic information corresponding to each group of video frames has been obtained as the supervision information in the first stage, it can be determined based on the reference optical flow characteristic information and the predicted optical flow characteristic information corresponding to each group of video frames Optical flow loss information.
  • the Euclidean distance between the reference optical flow characteristic information and the predicted optical flow characteristic information corresponding to each group of video frames may be determined as the optical flow loss information corresponding to each group of video frames.
  • the calculation formula of Euclidean distance can be seen in Equation 2.
  • loss flow is Euclidean distance
  • #feat is the number of groups of multiple video frames
  • multiple sets of video frames are input to the trained 3D convolutional neural module to obtain reference space feature information corresponding to each set of video frames, and the predicted optical flow feature information and reference space feature information corresponding to each set of video frames are obtained.
  • Perform feature fusion input the predicted optical flow feature information and reference space feature information corresponding to each group of video frames after fusion into the first classifier module, and determine the predicted classification category information corresponding to each group of video frames.
  • the classification loss information corresponding to each group of video frames is determined.
  • the cross-entropy distance between the standard classification category information corresponding to each group of video frames and the predicted classification category information may be calculated as the classification loss information corresponding to each group of video frames.
  • the weight parameters in the optical flow substitution module can be adjusted, and based on the classification loss information corresponding to each group of video frames, the weight in the first classifier module can be adjusted. Adjust the parameters.
  • the step of adjusting the weight parameter in the optical flow substitution module may include: based on the optical flow corresponding to each group of video frames The loss information, the classification loss information and the preset adjustment scale factor adjust the weight parameters in the optical flow replacement module.
  • the adjustment scale factor represents the adjustment range in the process of adjusting the weight parameter in the optical flow substitution module based on the optical flow loss information.
  • the weight parameters in the optical flow substitution module are affected by two aspects of loss information, that is, the optical flow loss information and classification loss information corresponding to each group of video frames, you can adjust the correspondence between each group of video frames by adjusting the scale factor
  • the optical flow loss information and classification loss information the adjustment range in the process of adjusting the weight parameters in the optical flow replacement module.
  • the calculation formula of optical flow loss information and classification loss information can be seen in Equation 3.
  • ⁇ is the adjustment scale factor
  • loss flow is the Euclidean distance
  • #feat is the number of groups of multiple video frames
  • It is the reference optical flow characteristic information corresponding to the i-th video frame.
  • the weight parameters in the optical flow substitution module can be adjusted by Equation 3 until the optical flow substitution module is determined to converge, and the trained optical flow substitution module is obtained. At this time, it can be considered that the optimized video action classification model has been trained. The running code corresponding to the stream module is deleted.
  • multiple video frames of the video to be classified can be directly input into the trained optimized video action classification model, and the trained optimized video action classification model can automatically classify the classified video, and Finally, the classification category information corresponding to the video to be classified is obtained, and the efficiency of classification processing is improved.
  • the video frame is directly used as the input of the optical flow substitution module in the model.
  • the optical flow substitution module can directly extract the optical flow characteristic information corresponding to multiple video frames of the video to be classified, and determine the classification category information corresponding to the video to be classified based on the optical flow characteristic information To further improve the efficiency of classification processing.
  • Fig. 5 is a block diagram of a device for classifying video actions according to an exemplary embodiment. 5, the device includes a first determination unit 510, a first input unit 520 and a second determination unit 530.
  • the first determining unit 510 is configured to obtain videos to be classified and determine multiple video frames in the videos to be classified;
  • the first input unit 520 is configured to input multiple video frames into the optical flow substitution module in the trained optimized video action classification model to obtain optical flow feature information corresponding to multiple video frames; input multiple video frames Go to the three-dimensional convolutional neural module in the optimized video action classification model after training to obtain the spatial feature information corresponding to multiple video frames;
  • the second determining unit 530 is configured to determine classification category information corresponding to the video to be classified based on the optical flow feature information and the spatial feature information.
  • the apparatus for classifying video actions further includes:
  • the first training unit is configured to train the video action classification model based on the training sample, where the training sample includes multiple sets of video frames and standard classification category information corresponding to each set of video frames, and the video action classification model includes a three-dimensional convolutional nerve Module and optical flow module;
  • the second input unit is configured to input multiple sets of video frames to the trained optical flow module to determine reference optical flow characteristic information corresponding to each set of video frames;
  • the establishment unit is configured to establish an optimized video motion classification model based on the trained three-dimensional convolutional neural module, the preset optical flow replacement module, and the preset classifier module;
  • the second training unit is configured to train the optimized video action classification model based on multiple sets of video frames, standard classification category information corresponding to each set of video frames, and reference optical flow feature information, to obtain the trained optimized video action classification model.
  • the second training unit is configured to:
  • the weight parameters in the optical flow substitution module are adjusted, and based on the classification loss information corresponding to each group of video frames, the weight parameters in the classifier module are adjusted.
  • the second training unit is configured to:
  • Adjust the weight parameters in the optical flow substitution module based on the optical flow loss information, classification loss information, and preset adjustment scale factor corresponding to each group of video frames, where adjusting the scale factor indicates optical flow substitution based on optical flow loss information The adjustment range during the adjustment of the weight parameters in the module.
  • the second training unit is configured to:
  • the Euclidean distance between the reference optical flow characteristic information and the predicted optical flow characteristic information corresponding to each group of video frames is determined as the optical flow loss information corresponding to each group of video frames.
  • multiple video frames of the video to be classified can be directly input into the trained optimized video action classification model, and the trained optimized video action classification model can automatically classify the classified video, and Finally, the classification category information corresponding to the video to be classified is obtained, and the efficiency of classification processing is improved.
  • the video frame is directly used as the input of the optical flow substitution module in the model.
  • the optical flow substitution module can directly extract the optical flow characteristic information corresponding to multiple video frames of the video to be classified, and determine the classification category information corresponding to the video to be classified based on the optical flow characteristic information To further improve the efficiency of classification processing.
  • Fig. 6 is a block diagram of a device 600 for video action classification according to an exemplary embodiment.
  • the apparatus 600 may be the computer equipment provided by the embodiments of the present application.
  • the device 600 may include one or more of the following components: a processing component 602, a memory 604, a power component 606, a multimedia component 608, an audio component 610, an input/output (I/O) interface 612, a sensor component 614, ⁇ 616.
  • the processing component 602 generally controls the overall operations of the device 600, such as operations associated with display, data communication, and recording operations.
  • the processing component 602 may include one or more processors 620 to execute instructions to complete all or part of the steps in the above method.
  • the processing component 602 may include one or more modules to facilitate interaction between the processing component 602 and other components.
  • the processing component 602 may include a multimedia module to facilitate interaction between the multimedia component 608 and the processing component 602.
  • the memory 604 is configured to store various types of data to support operation at the device 600. Examples of these data include instructions, messages, pictures, videos, etc. for any applications or methods operating on the device 600.
  • the memory 604 may be implemented by any type of volatile or nonvolatile storage device or a combination thereof, such as static random access memory (SRAM), electrically erasable programmable read only memory (EEPROM), erasable and removable Programmable read only memory (EPROM), programmable read only memory (PROM), read only memory (ROM), magnetic memory, flash memory, magnetic disk or optical disk.
  • SRAM static random access memory
  • EEPROM electrically erasable programmable read only memory
  • EPROM erasable and removable Programmable read only memory
  • PROM programmable read only memory
  • ROM read only memory
  • magnetic memory flash memory
  • flash memory magnetic disk or optical disk.
  • the power supply component 606 provides power to various components of the device 600.
  • the power component 606 may include a power management system, one or more power sources, and other components associated with generating, managing, and distributing power for the device 600.
  • the multimedia component 608 includes a screen between the device 600 and the user that provides an output interface.
  • the screen may include a liquid crystal display (LCD) and a touch panel (TP). If the screen includes a touch panel, the screen may be implemented as a touch screen to receive input signals from the user.
  • LCD liquid crystal display
  • TP touch panel
  • the audio component 610 is configured to output and/or input audio signals.
  • the audio component 610 includes a microphone (MIC).
  • the microphone When the device 600 is in an operation mode, such as a recording mode and a voice recognition mode, the microphone is configured to receive an external audio signal.
  • the received audio signal may be further stored in the memory 604 or transmitted via the communication component 616.
  • the audio component 610 further includes a speaker for outputting audio signals.
  • the I/O interface 612 provides an interface between the processing component 602 and a peripheral interface module.
  • the peripheral interface module may be a keyboard, a click wheel, or a button. These buttons may include, but are not limited to: home button, volume button, start button, and lock button.
  • the sensor component 614 includes one or more sensors for providing the device 600 with status assessments in various aspects.
  • the sensor component 614 can detect the on/off state of the device 600, the relative positioning of the components, such as the display and keypad of the device 600 and the temperature change of the device 600.
  • the communication component 616 is configured to facilitate wired or wireless communication between the device 600 and other devices.
  • the device 600 may access a wireless network based on a communication standard, such as WiFi, an operator network (such as 2G, 3G, 4G, or 5G), or a combination thereof.
  • a communication standard such as WiFi, an operator network (such as 2G, 3G, 4G, or 5G), or a combination thereof.
  • the communication component 616 receives a broadcast signal or broadcast related information from an external broadcast management system via a broadcast channel.
  • the apparatus 600 may be one or more application specific integrated circuits (ASICs), digital signal processors (DSPs), digital signal processing devices (DSPDs), programmable logic devices (PLDs), field programmable A gate array (FPGA), controller, microcontroller, microprocessor or other electronic components are implemented to perform the above method.
  • ASICs application specific integrated circuits
  • DSPs digital signal processors
  • DSPDs digital signal processing devices
  • PLDs programmable logic devices
  • FPGA field programmable A gate array
  • controller microcontroller, microprocessor or other electronic components are implemented to perform the above method.
  • a non-transitory computer-readable storage medium including instructions is also provided, for example, a memory 604 including instructions, which can be executed by the processor 620 of the device 600 to complete the above method.
  • the non-transitory computer-readable storage medium may be ROM, random access memory (RAM), CD-ROM, magnetic tape, floppy disk, optical data storage device, or the like.
  • a computer program product is also provided, which when the computer program product is executed by the processor 620 of the device 600, enables the device 600 to execute to complete the above method.

Landscapes

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

Abstract

本申请是关于一种视频动作分类的方法、装置、计算机设备和存储介质，涉及机器学习模型技术领域。所述方法包括：获取待分类视频，确定待分类视频中的多个视频帧；将多个视频帧输入到训练后的优化视频动作分类模型中的光流替代模块中，得到多个视频帧对应的光流特征信息；将多个视频帧输入到训练后的优化视频动作分类模型中的三维卷积神经模块中，得到多个视频帧对应的空间特征信息；基于光流特征信息和空间特征信息，确定待分类视频对应的分类类别信息。采用本申请，可以将待分类视频的多个视频帧直接作为模型中的光流替代模块的输入，光流替代模块可以直接提取待分类视频的多个视频帧对应的光流特征信息，进一步提高了分类处理的效率。

Description

视频动作分类的方法、装置、计算机设备和存储介质
相关申请的交叉引用
本申请要求在2018年11月28日提交中国专利局、申请号为201811437221.X、发明名称为“视频动作分类的方法、装置、计算机设备和存储介质”的中国专利申请的优先权，其全部内容通过引用结合在本申请中。
技术领域
本申请是关于机器学习模型技术领域，尤其是关于一种视频动作分类的方法、装置、计算机设备和存储介质。
背景技术
随着社会的发展，越来越多的用户喜欢利用碎片时间观看或者拍摄短视频。当任一用户将拍摄的短视频上传到短视频平台时，短视频平台中的相关人员可以查看短视频，并根据主观意念对短视频中的对象的动作进行分类，如跳舞、爬树、喝水等，然后相关人员可以根据分类结果为短视频添加对应的标签。
在实现本申请的过程中，发明人发现至少存在以下问题：
由于短视频平台接收到的短视频的数量巨大，如果通过人工的方式为每一个短视频中的对象的动作进行分类，会导致分类操作的效率极低。
发明内容
为克服相关技术中短视频平台接收到的短视频数量巨大，通过人工方式分类会导致分类操作效率极低的问题，本申请实施例提供一种视频动作分类的方法和装置。
根据本申请实施例的第一方面，提供一种视频动作分类的方法，包括：
获取待分类视频，确定待分类视频中的多个视频帧；
将多个视频帧输入到训练后的优化视频动作分类模型中的光流替代模块中，得到多个视频帧对应的光流特征信息；
将多个视频帧输入到训练后的优化视频动作分类模型中的三维卷积神经模块中，得到多个视频帧对应的空间特征信息；
基于光流特征信息和空间特征信息，确定待分类视频对应的分类类别信息。
根据本申请实施例的第二方面，提供一种视频动作分类的装置，包括：
第一确定单元，被配置为获取待分类视频，确定所述待分类视频中的多个视频帧；
第一输入单元，被配置为将所述多个视频帧输入到训练后的优化视频动作分类模型中的光流替代模块中，得到所述多个视频帧对应的光流特征信息；将所述多个视频帧输入到所述训练后的优化视频动作分类模型中的三维卷积神经模块中，得到所述多个视频帧对应的空间特征信息；
第二确定单元，被配置为基于所述光流特征信息和所述空间特征信息，确定所述待分类视频对应的分类类别信息。
根据本申请实施例的第三方面，提供一种计算机设备，包括：
处理器；
用于存储处理器可执行指令的存储器；
其中，所述处理器被配置为：
获取待分类视频，确定所述待分类视频中的多个视频帧；
将所述多个视频帧输入到训练后的优化视频动作分类模型中的光流替代模块中，得到所述多个视频帧对应的光流特征信息；
将所述多个视频帧输入到所述训练后的优化视频动作分类模型中的三维卷积神经模块中，得到所述多个视频帧对应的空间特征信息；
基于所述光流特征信息和所述空间特征信息，确定所述待分类视频对应的分类类别信息。
根据本申请实施例的第四方面，提供一种非临时性计算机可读存储介质，当所述存储介质中的指令由计算机设备的处理器执行时，使得所述计算机设备能够执行一种视频动作分类的方法，所述方法包括：
获取待分类视频，确定所述待分类视频中的多个视频帧；
将所述多个视频帧输入到训练后的优化视频动作分类模型中的光流替代模块中，得到所述多个视频帧对应的光流特征信息；
将所述多个视频帧输入到所述训练后的优化视频动作分类模型中的三维卷积神经模块中，得到所述多个视频帧对应的空间特征信息；
基于所述光流特征信息和所述空间特征信息，确定所述待分类视频对应的分类类别信息。
根据本申请实施例的第五方面，提供一种计算机程序产品，当所述计算机程序产品由计算机设备的处理器执行时，使得所述计算机设备能够执行一种视频动作分类的方法，所述方法包括：
获取待分类视频，确定所述待分类视频中的多个视频帧；
将所述多个视频帧输入到训练后的优化视频动作分类模型中的光流替代模块中，得到所述多个视频帧对应的光流特征信息；
将所述多个视频帧输入到所述训练后的优化视频动作分类模型中的三维卷积神经模块中，得到所述多个视频帧对应的空间特征信息；
基于所述光流特征信息和所述空间特征信息，确定所述待分类视频对应的分类类别信息。
通过本申请实施例提供的方法，可以将待分类视频的多个视频帧直接输入到训练后的优化视频动作分类模型中，训练后的优化视频动作分类模型可以自动对待分类视频进行分类处理，并最终得到待分类视频对应的分类类别信息，提高了分类处理的效率。在训练后的优化视频动作分类模型对待分类视频进行分类处理的过程中，无需再预先基于待分类视频的多个视频帧确定多个视频帧对应的光流图，可以将待分类视频的多个视频帧直接作为模型中的光流替代模块的输入，光流替代模块可以直接提取待分类视频的多个视频 帧对应的光流特征信息，基于光流特征信息确定待分类视频对应的分类类别信息，进一步提高了分类处理的效率。
附图说明
此处的附图被并入说明书中并构成本说明书的一部分，示出了符合本申请实施例，并与说明书一起用于解释本申请的原理。
图1是根据一示例性实施例示出的一种视频动作分类的方法的流程图；
图2是根据一示例性实施例示出的一种视频动作分类的方法的流程图；
图3是根据一示例性实施例示出的一种训练优化视频动作分类模型的方法的流程图；
图4是根据一示例性实施例示出的一种训练优化视频动作分类模型的方法的流程图；
图5是根据一示例性实施例示出的一种视频动作分类的装置的框图；
图6是根据一示例性实施例示出的一种用于视频动作分类的装置的框图。
具体实施方式
这里将详细地对示例性实施例进行说明，其示例表示在附图中。下面的描述涉及附图时，除非另有表示，不同附图中的相同数字表示相同或相似的要素。以下示例性实施例中所描述的实施方式并不代表与本申请相一致的所有实施方式。相反，它们仅是与如所附权利要求书中所详述的、本申请的一些方面相一致的装置和方法的例子。
随着社会的发展，越来越多的用户喜欢利用碎片时间观看或者拍摄短视频。当用户将拍摄的短视频上传到短视频平台时，短视频平台需要对短视频中的对象的动作进行分类，如跳舞、爬树、喝水等，然后根据分类结果为短视频添加对应的标签。在本申请实施例中，提供可以自动为短视频进行分类处理的方法。
图1是根据一示例性实施例示出的一种视频动作分类的方法的流程图， 如图1所示，视频动作分类的方法用于短视频平台的服务器中，包括以下步骤。
在步骤S110中，获取待分类视频，确定待分类视频中的多个视频帧。
在实施中，短视频平台的服务器可以接收到用户上传的大量的短视频，任一短视频都可以作为待分类视频，因此服务器可以获取到待分类视频。由于一条待分类视频是由许多个视频帧构成的，无需将一条待分类视频中的所有的视频帧都用于后续步骤，因此服务器可以在一条待分类视频中的所有的视频帧中，提取预设数目的多个视频帧。可选地，服务器可以在一条待分类视频中的所有的视频帧中，随机提取预设数目的多个视频帧。其中，预设数目可以根据经验值进行设定，例如，预设数目为10，预设数目为5等。
在步骤S120中，将多个视频帧输入到训练后的优化视频动作分类模型中的光流替代模块中，得到多个视频帧对应的光流特征信息。
在实施中，可以预先训练优化视频动作分类模型，用于对待分类视频进行分类处理。优化视频动作分类模型包括多个功能模块，每个功能模块都有不同的作用。优化视频动作分类模型可以包括光流替代模块、三维卷积神经模块以及第一分类器模块。
光流替代模块用于提取多个视频帧对应的光流特征信息。如图2所示，当服务器将多个视频帧输入到训练后的优化视频动作分类模型中的光流替代模块中时，光流替代模块可以输出多个视频帧对应的光流特征信息。其中，光流特征信息表示多个视频帧中包括的对象对应的运动矢量，即对象在多个视频帧中的最先拍摄的视频帧中的位置是朝着什么样的方向运动至最后拍摄的视频帧中的位置的。
在步骤S130中，将多个视频帧输入到训练后的优化视频动作分类模型中的三维卷积神经模块中，得到多个视频帧对应的空间特征信息。
其中，三维卷积神经模块可以包括C3D(3 Dimensions Convolution，三维卷积)模块。
在实施中，三维卷积神经模块用于提取多个视频帧对应的空间特征信息。 如图2所示，当服务器将多个视频帧输入到训练后的优化视频动作分类模型中的三维卷积神经模块中时，三维卷积神经模块可以输出多个视频帧对应的空间特征信息。其中，空间特征信息表示多个视频帧中包括的对象在每个视频帧中的位置，空间特征信息可以由一组三维信息构成，三维信息中的二维可以表示对象在一个视频帧中的位置，最后一维可以表示该视频帧对应的拍摄时刻。
在步骤S140中，基于光流特征信息和空间特征信息，确定待分类视频对应的分类类别信息。
在实施中，在得到多个视频帧对应的光流特征信息和空间特征信息之后，服务器可以将光流特征信息和空间特征信息进行特征融合。具体地，可以通过CONCAT语句将光流特征信息和空间特征信息进行特征融合，将融合后的光流特征信息和空间特征信息，输入到第一分类器模块中，然后可以由第一分类器模块输出光流特征信息和空间特征信息对应的分类类别信息，作为待分类视频对应的分类类别信息，实现了端到端的分类处理。
在一种可能的实现方式中，如图3所示，本申请实施例提供的方法还可以包括：
在步骤S310中，基于训练样本，对视频动作分类模型进行训练，其中，训练样本包括多组视频帧以及每组视频帧对应的标准分类类别信息，视频动作分类模型包括三维卷积神经模块和光流模块。
在步骤S320中，将多组视频帧分别输入到训练后的光流模块，确定每组视频帧对应的参考光流特征信息。
在步骤S330中，基于训练后的三维卷积神经模块、预设的光流替代模块和第一分类器模块，建立优化视频动作分类模型。
在步骤S340中，基于多组视频帧、每组视频帧对应的标准分类类别信息和参考光流特征信息，对优化视频动作分类模型进行训练，得到训练后的优化视频动作分类模型。
在实施中，在使用训练后的优化视频动作分类模型对待分类视频进行分 类处理之前，需要预先训练好优化视频动作分类模型。在本申请实施例中，训练优化视频动作分类模型的过程可以分为两个阶段。第一个阶段，可以基于训练样本，对视频动作分类模型进行训练。第二个阶段，可以将多组视频帧分别输入到训练后的光流模块，确定每组视频帧对应的参考光流特征信息，基于训练后的三维卷积神经模块、预设的光流替代模块和第一分类器模块，建立优化视频动作分类模型，基于多组视频帧、每组视频帧对应的标准分类类别信息和参考光流特征信息，对优化视频动作分类模型进行训练，得到训练后的优化视频动作分类模型。
如图4所示，在第一个阶段，首先可以基于三维卷积神经模块、光流模块和第二分类器模块，建立视频动作分类模型。其中，三维卷积神经模块用于提取一组视频帧对应的空间特征信息，光流模块用于提取一组视频帧对应的光流特征信息，第二分类器模块用于基于一组视频帧对应的空间特征信息和光流特征信息，确定一组视频帧对应的预测分类类别信息。
在实施中，将训练样本中的多组视频帧分别输入到三维卷积神经模块中，三维卷积神经模块可以提取每组视频帧对应的空间特征信息，同时可以不通过视频动作分类模型，预先基于多组视频帧，分别确定每组视频帧对应的光流图，将每组视频帧对应的光流图输入到光流模块中，光流模块可以输出每组视频帧对应的光流特征信息。然后可以将每组视频帧对应的空间特征信息和光流特征信息进行特征融合，将融合后的每组视频帧对应的空间特征信息和光流特征信息输入到第二分类器模块中，第二分类器模块可以输出每组视频帧对应的预测分类类别信息。
在实施中，将训练样本中的每组视频帧对应的标准分类类别信息作为监督信息，确定每组视频帧对应的预测分类类别信息和标准分类类别信息之间的差值信息。接着可以基于每组视频帧对应的差值信息，对视频动作分类模型中的权重参数进行调整。随后可以重复执行上述过程，直到确定视频动作分类模型收敛，得到训练后的视频动作分类模型。其中，差值信息可以是交叉熵距离。交叉熵距离的计算公式可以见公式1。
Figure PCTCN2019106250-appb-000001
其中，loss entropy为交叉熵距离，
Figure PCTCN2019106250-appb-000002
为预测分类类别信息，y为标准分类类别信息。
如图4所示，在第二个阶段，由于在第一阶段已经将视频动作分类模型训练好，视频动作分类模型中的光流模块也已训练好，即训练后的光流模块可以准确提取每组视频帧对应的光流特征信息。因此，可以将收敛后的光流模块输出的参考光流特征信息作为监督信息，添加到训练样本中，用于后续对其他模块的训练。
在检测到光流模块收敛时，可以冻结光流模块中的权重参数，不再继续对光流模块中的权重参数进行调整。然后可以将三维卷积神经模块、预设的光流替代模块和第一分类器模块，作为优化视频动作分类模型中的模块，对优化视频动作分类模型进行训练。
在一种可能的实施方式中，可以继续对三维卷积神经模块进行训练，使得三维卷积神经模块输出的结果的精确度越来越高，同时还可以对光流替代模块进行训练，使得光流替代模块可以替代光流模块提取每组视频帧对应的光流特征信息。
在一种可能的实施方式中，可以基于多组视频帧、每组视频帧对应的标准分类类别信息和参考光流特征信息，对优化视频动作分类模型进行训练，得到训练后的优化视频动作分类模型。
在一种可能的实现方式中，步骤S340可以包括：将多组视频帧分别输入到光流替代模块，得到每组视频帧对应的预测光流特征信息；基于每组视频帧对应的参考光流特征信息和预测光流特征信息，确定每组视频帧对应的光流损失信息；将多组视频帧分别输入到训练后的三维卷积神经模块，得到每组视频帧对应的参考空间特征信息；将每组视频帧对应的预测光流特征信息和参考空间特征信息，输入到第一分类器模块，确定每组视频帧对应的预测分类类别信息；基于每组视频帧对应的标准分类类别信息和预测分类类别信 息，确定每组视频帧对应的分类损失信息；基于每组视频帧对应的光流损失信息和分类损失信息，对光流替代模块中的权重参数进行调整，基于每组视频帧对应的分类损失信息，对第一分类器模块中的权重参数进行调整。
在实施中，可以直接将多组视频帧分别输入到光流替代模块中，无需预先在优化视频动作分类模型之外，单独基于多组视频帧，分别确定每组视频帧对应的光流图。光流替代模块可以直接将多组视频帧分别作为输入，而无需将光流图作为输入。当将多组视频帧分别输入到光流替代模块中时，光流替代模块可以输出每组视频帧对应的预测光流特征信息。
由于在第一阶段已经得到每组视频帧对应的参考光流特征信息，作为监督信息，因此可以基于每组视频帧对应的参考光流特征信息和预测光流特征信息，确定每组视频帧对应的光流损失信息。
在一种可能的实施方式中，可以确定每组视频帧对应的参考光流特征信息和预测光流特征信息之间的欧氏距离，作为每组视频帧对应的光流损失信息。欧式距离的计算公式可以见公式2。
Figure PCTCN2019106250-appb-000003
其中，loss flow为欧式距离，#feat为多组视频帧的组数，
Figure PCTCN2019106250-appb-000004
为第i组视频帧对应的预测光流特征信息，
Figure PCTCN2019106250-appb-000005
为第i组视频帧对应的参考光流特征信息。
在实施中，将多组视频帧分别输入到训练后的三维卷积神经模块，得到每组视频帧对应的参考空间特征信息，将每组视频帧对应的预测光流特征信息和参考空间特征信息进行特征融合，将融合后的每组视频帧对应的预测光流特征信息和参考空间特征信息输入到第一分类器模块，确定每组视频帧对应的预测分类类别信息。
在实施中，基于每组视频帧对应的标准分类类别信息和预测分类类别信息，确定每组视频帧对应的分类损失信息。
在一种可能的实施方式中，可以计算每组视频帧对应的标准分类类别信 息和预测分类类别信息之间的交叉熵距离，作为每组视频帧对应的分类损失信息。最后可以基于每组视频帧对应的光流损失信息和分类损失信息，对光流替代模块中的权重参数进行调整，基于每组视频帧对应的分类损失信息，对第一分类器模块中的权重参数进行调整。
在一种可能的实现方式中，基于每组视频帧对应的光流损失信息和分类损失信息，对光流替代模块中的权重参数进行调整的步骤可以包括：基于每组视频帧对应的光流损失信息、分类损失信息和预设的调整比例系数，对光流替代模块中的权重参数进行调整。
其中，调整比例系数表示基于光流损失信息对光流替代模块中的权重参数进行调整的过程中的调整幅度。
在实施中，由于光流替代模块中的权重参数受两方面的损失信息的影响，即每组视频帧对应的光流损失信息和分类损失信息，因此可以通过调整比例系数调整每组视频帧对应的光流损失信息和分类损失信息，对光流替代模块中的权重参数进行调整的过程中的调整幅度。光流损失信息和分类损失信息的计算公式可见公式3。
Figure PCTCN2019106250-appb-000006
其中，
Figure PCTCN2019106250-appb-000007
为分类损失信息，λ为调整比例系数，loss flow为欧式距离，#feat为多组视频帧的组数，
Figure PCTCN2019106250-appb-000008
为第i组视频帧对应的预测光流特征信息，
Figure PCTCN2019106250-appb-000009
为第i组视频帧对应的参考光流特征信息。
可以通过公式3对光流替代模块中的权重参数进行调整，直到确定光流替代模块收敛，得到训练后的光流替代模块，此时可以认为优化视频动作分类模型已经训练好了，可以将光流模块对应的运行代码进行删除。
通过本申请实施例提供的方法，可以将待分类视频的多个视频帧直接输入到训练后的优化视频动作分类模型中，训练后的优化视频动作分类模型可以自动对待分类视频进行分类处理，并最终得到待分类视频对应的分类类别信息，提高了分类处理的效率。在训练后的优化视频动作分类模型对待分类 视频进行分类处理的过程中，无需再预先基于待分类视频的多个视频帧确定多个视频帧对应的光流图，可以将待分类视频的多个视频帧直接作为模型中的光流替代模块的输入，光流替代模块可以直接提取待分类视频的多个视频帧对应的光流特征信息，基于光流特征信息确定待分类视频对应的分类类别信息，进一步提高了分类处理的效率。
图5是根据一示例性实施例示出的一种视频动作分类的装置框图。参照图5，该装置包括第一确定单元510，第一输入单元520和第二确定单元530。
第一确定单元510，被配置为获取待分类视频，确定待分类视频中的多个视频帧；
第一输入单元520，被配置为将多个视频帧输入到训练后的优化视频动作分类模型中的光流替代模块中，得到多个视频帧对应的光流特征信息；将多个视频帧输入到训练后的优化视频动作分类模型中的三维卷积神经模块中，得到多个视频帧对应的空间特征信息；
第二确定单元530，被配置为基于光流特征信息和空间特征信息，确定待分类视频对应的分类类别信息。
可选地，视频动作分类的装置还包括：
第一训练单元，被配置为基于训练样本，对视频动作分类模型进行训练，其中，训练样本包括多组视频帧以及每组视频帧对应的标准分类类别信息，视频动作分类模型包括三维卷积神经模块和光流模块；
第二输入单元，被配置为将多组视频帧分别输入到训练后的光流模块，确定每组视频帧对应的参考光流特征信息；
建立单元，被配置为基于训练后的三维卷积神经模块、预设的光流替代模块和预设的分类器模块，建立优化视频动作分类模型；
第二训练单元，被配置为基于多组视频帧、每组视频帧对应的标准分类类别信息和参考光流特征信息，对优化视频动作分类模型进行训练，得到训练后的优化视频动作分类模型。
可选地，第二训练单元被配置为：
将多组视频帧分别输入到光流替代模块，得到每组视频帧对应的预测光流特征信息；
基于每组视频帧对应的参考光流特征信息和预测光流特征信息，确定每组视频帧对应的光流损失信息；
将多组视频帧分别输入到训练后的三维卷积神经模块，得到每组视频帧对应的参考空间特征信息；
将每组视频帧对应的预测光流特征信息和参考空间特征信息，输入到分类器模块，确定每组视频帧对应的预测分类类别信息；
基于每组视频帧对应的标准分类类别信息和预测分类类别信息，确定每组视频帧对应的分类损失信息；
基于每组视频帧对应的光流损失信息和分类损失信息，对光流替代模块中的权重参数进行调整，基于每组视频帧对应的分类损失信息，对分类器模块中的权重参数进行调整。
可选地，第二训练单元被配置为：
基于每组视频帧对应的光流损失信息、分类损失信息和预设的调整比例系数，对光流替代模块中的权重参数进行调整，其中，调整比例系数表示基于光流损失信息对光流替代模块中的权重参数进行调整的过程中的调整幅度。
可选地，第二训练单元被配置为：
确定每组视频帧对应的参考光流特征信息和预测光流特征信息之间的欧氏距离，作为每组视频帧对应的光流损失信息。
通过本申请实施例提供的装置，可以将待分类视频的多个视频帧直接输入到训练后的优化视频动作分类模型中，训练后的优化视频动作分类模型可以自动对待分类视频进行分类处理，并最终得到待分类视频对应的分类类别信息，提高了分类处理的效率。在训练后的优化视频动作分类模型对待分类视频进行分类处理的过程中，无需再预先基于待分类视频的多个视频帧确定多个视频帧对应的光流图，可以将待分类视频的多个视频帧直接作为模型中的光流替代模块的输入，光流替代模块可以直接提取待分类视频的多个视频 帧对应的光流特征信息，基于光流特征信息确定待分类视频对应的分类类别信息，进一步提高了分类处理的效率。
关于上述实施例中的装置，其中各个模块执行操作的具体方式已经在有关该方法的实施例中进行了详细描述，此处将不做详细阐述说明。
图6是根据一示例性实施例示出的一种用于视频动作分类的装置600的框图。例如，装置600可以是本申请实施例提供的计算机设备。
参照图6，装置600可以包括以下一个或多个组件：处理组件602，存储器604，电源组件606，多媒体组件608，音频组件610，输入/输出(I/O)的接口612，传感器组件614，以及通信组件616。
处理组件602通常控制装置600的整体操作，诸如与显示，数据通信，和记录操作相关联的操作。处理组件602可以包括一个或多个处理器620来执行指令，以完成上述的方法的全部或部分步骤。此外，处理组件602可以包括一个或多个模块，便于处理组件602和其他组件之间的交互。例如，处理组件602可以包括多媒体模块，以方便多媒体组件608和处理组件602之间的交互。
存储器604被配置为存储各种类型的数据以支持在装置600的操作。这些数据的示例包括用于在装置600上操作的任何应用程序或方法的指令，消息，图片，视频等。存储器604可以由任何类型的易失性或非易失性存储设备或者它们的组合实现，如静态随机存取存储器(SRAM)，电可擦除可编程只读存储器(EEPROM)，可擦除可编程只读存储器(EPROM)，可编程只读存储器(PROM)，只读存储器(ROM)，磁存储器，快闪存储器，磁盘或光盘。
电源组件606为装置600的各种组件提供电力。电源组件606可以包括电源管理系统，一个或多个电源，及其他与为装置600生成、管理和分配电力相关联的组件。
多媒体组件608包括在装置600和用户之间的提供一个输出接口的屏幕。在一些实施例中，屏幕可以包括液晶显示器(LCD)和触摸面板(TP)。如果 屏幕包括触摸面板，屏幕可以被实现为触摸屏，以接收来自用户的输入信号。
音频组件610被配置为输出和/或输入音频信号。例如，音频组件610包括一个麦克风(MIC)，当装置600处于操作模式，如记录模式和语音识别模式时，麦克风被配置为接收外部音频信号。所接收的音频信号可以被进一步存储在存储器604或经由通信组件616发送。在一些实施例中，音频组件610还包括一个扬声器，用于输出音频信号。
I/O接口612为处理组件602和外围接口模块之间提供接口，上述外围接口模块可以是键盘，点击轮，按钮等。这些按钮可包括但不限于：主页按钮、音量按钮、启动按钮和锁定按钮。
传感器组件614包括一个或多个传感器，用于为装置600提供各个方面的状态评估。例如，传感器组件614可以检测到设备600的打开/关闭状态，组件的相对定位，例如组件为装置600的显示器和小键盘和装置600的温度变化。
通信组件616被配置为便于装置600和其他设备之间有线或无线方式的通信。装置600可以接入基于通信标准的无线网络，如WiFi，运营商网络(如2G、3G、4G或5G)，或它们的组合。在一个示例性实施例中，通信组件616经由广播信道接收来自外部广播管理系统的广播信号或广播相关信息。
在示例性实施例中，装置600可以被一个或多个应用专用集成电路(ASIC)、数字信号处理器(DSP)、数字信号处理设备(DSPD)、可编程逻辑器件(PLD)、现场可编程门阵列(FPGA)、控制器、微控制器、微处理器或其他电子元件实现，用于执行上述方法。
在示例性实施例中，还提供了一种包括指令的非临时性计算机可读存储介质，例如包括指令的存储器604，上述指令可由装置600的处理器620执行以完成上述方法。例如，非临时性计算机可读存储介质可以是ROM、随机存取存储器(RAM)、CD-ROM、磁带、软盘和光数据存储设备等。
在示例性实施例中，还提供了一种计算机程序产品，当计算机程序产品由装置600的处理器620执行时，使得装置600能够执行以完成上述方法。
本领域技术人员在考虑说明书及实践这里公开的发明后，将容易想到本申请的其它实施方案。本申请旨在涵盖本申请的任何变型、用途或者适应性变化，这些变型、用途或者适应性变化遵循本申请的一般性原理并包括本申请未公开的本技术领域中的公知常识或惯用技术手段。说明书和实施例仅被视为示例性的，本申请的真正范围和精神由下面的权利要求指出。
应当理解的是，本申请并不局限于上面已经描述并在附图中示出的精确结构，并且可以在不脱离其范围进行各种修改和改变。本申请的范围仅由所附的权利要求来限制。

Claims (12)

  1. 一种视频动作分类的方法，包括：
    获取待分类视频，确定所述待分类视频中的多个视频帧；
    将所述多个视频帧输入到训练后的优化视频动作分类模型中的光流替代模块中，得到所述多个视频帧对应的光流特征信息；
    将所述多个视频帧输入到所述训练后的优化视频动作分类模型中的三维卷积神经模块中，得到所述多个视频帧对应的空间特征信息；
    基于所述光流特征信息和所述空间特征信息，确定所述待分类视频对应的分类类别信息。
  2. 根据权利要求1所述的方法，所述方法还包括：
    基于训练样本，对视频动作分类模型进行训练，其中，所述训练样本包括多组视频帧以及每组视频帧对应的标准分类类别信息，所述视频动作分类模型包括三维卷积神经模块和光流模块；
    将所述多组视频帧分别输入到训练后的光流模块，确定每组视频帧对应的参考光流特征信息；
    基于训练后的三维卷积神经模块、预设的光流替代模块和预设的第一分类器模块，建立优化视频动作分类模型；
    基于所述多组视频帧、每组视频帧对应的标准分类类别信息和参考光流特征信息，对所述优化视频动作分类模型进行训练，得到训练后的优化视频动作分类模型。
  3. 根据权利要求2所述的方法，所述基于所述多组视频帧、每组视频帧对应的标准分类类别信息和参考光流特征信息，对所述优化视频动作分类模型进行训练，包括：
    将所述多组视频帧分别输入到所述光流替代模块，得到每组视频帧对应的预测光流特征信息；
    基于每组视频帧对应的参考光流特征信息和预测光流特征信息，确定每 组视频帧对应的光流损失信息；
    将所述多组视频帧分别输入到训练后的三维卷积神经模块，得到每组视频帧对应的参考空间特征信息；
    将每组视频帧对应的预测光流特征信息和参考空间特征信息，输入到预设的第二分类器模块，确定每组视频帧对应的预测分类类别信息；
    基于每组视频帧对应的标准分类类别信息和预测分类类别信息，确定每组视频帧对应的分类损失信息；
    基于每组视频帧对应的光流损失信息和分类损失信息，对所述光流替代模块中的权重参数进行调整，基于每组视频帧对应的分类损失信息，对所述第一分类器模块中的权重参数进行调整。
  4. 根据权利要求3所述的方法，所述基于每组视频帧对应的光流损失信息和分类损失信息，对所述光流替代模块中的权重参数进行调整，包括：
    基于每组视频帧对应的光流损失信息、分类损失信息和预设的调整比例系数，对所述光流替代模块中的权重参数进行调整，其中，所述调整比例系数表示基于光流损失信息对所述光流替代模块中的权重参数进行调整的过程中的调整幅度。
  5. 根据权利要求3所述的方法，所述基于每组视频帧对应的参考光流特征信息和预测光流特征信息，确定每组视频帧对应的光流损失信息，包括：
    确定每组视频帧对应的参考光流特征信息和预测光流特征信息之间的欧氏距离，作为每组视频帧对应的光流损失信息。
  6. 一种视频动作分类的装置，包括：
    第一确定单元，被配置为获取待分类视频，确定所述待分类视频中的多个视频帧；
    第一输入单元，被配置为将所述多个视频帧输入到训练后的优化视频动作分类模型中的光流替代模块中，得到所述多个视频帧对应的光流特征信息；将所述多个视频帧输入到所述训练后的优化视频动作分类模型中的三维卷积神经模块中，得到所述多个视频帧对应的空间特征信息；
    第二确定单元，被配置为基于所述光流特征信息和所述空间特征信息，确定所述待分类视频对应的分类类别信息。
  7. 一种计算机设备，包括：
    处理器；
    用于存储处理器可执行指令的存储器；
    其中，所述处理器被配置为：
    获取待分类视频，确定所述待分类视频中的多个视频帧；
    将所述多个视频帧输入到训练后的优化视频动作分类模型中的光流替代模块中，得到所述多个视频帧对应的光流特征信息；
    将所述多个视频帧输入到所述训练后的优化视频动作分类模型中的三维卷积神经模块中，得到所述多个视频帧对应的空间特征信息；
    基于所述光流特征信息和所述空间特征信息，确定所述待分类视频对应的分类类别信息。
  8. 根据权利要求7所述的计算机设备，包括：
    基于训练样本，对视频动作分类模型进行训练，其中，所述训练样本包括多组视频帧以及每组视频帧对应的标准分类类别信息，所述视频动作分类模型包括三维卷积神经模块和光流模块；
    将所述多组视频帧分别输入到训练后的光流模块，确定每组视频帧对应的参考光流特征信息；
    基于训练后的三维卷积神经模块、预设的光流替代模块和预设的第一分类器模块，建立优化视频动作分类模型；
    基于所述多组视频帧、每组视频帧对应的标准分类类别信息和参考光流特征信息，对所述优化视频动作分类模型进行训练，得到训练后的优化视频动作分类模型。
  9. 根据权利要求8所述的计算机设备，所述基于所述多组视频帧、每组视频帧对应的标准分类类别信息和参考光流特征信息，对所述优化视频动作分类模型进行训练，包括：
    将所述多组视频帧分别输入到所述光流替代模块，得到每组视频帧对应的预测光流特征信息；
    基于每组视频帧对应的参考光流特征信息和预测光流特征信息，确定每组视频帧对应的光流损失信息；
    将所述多组视频帧分别输入到训练后的三维卷积神经模块，得到每组视频帧对应的参考空间特征信息；
    将每组视频帧对应的预测光流特征信息和参考空间特征信息，输入到预设的第二分类器模块，确定每组视频帧对应的预测分类类别信息；
    基于每组视频帧对应的标准分类类别信息和预测分类类别信息，确定每组视频帧对应的分类损失信息；
    基于每组视频帧对应的光流损失信息和分类损失信息，对所述光流替代模块中的权重参数进行调整，基于每组视频帧对应的分类损失信息，对所述第一分类器模块中的权重参数进行调整。
  10. 根据权利要求9所述的计算机设备，所述基于每组视频帧对应的光流损失信息和分类损失信息，对所述光流替代模块中的权重参数进行调整，包括：
    基于每组视频帧对应的光流损失信息、分类损失信息和预设的调整比例系数，对所述光流替代模块中的权重参数进行调整，其中，所述调整比例系数表示基于光流损失信息对所述光流替代模块中的权重参数进行调整的过程中的调整幅度。
  11. 根据权利要求9所述的计算机设备，所述基于每组视频帧对应的参考光流特征信息和预测光流特征信息，确定每组视频帧对应的光流损失信息，包括：
    确定每组视频帧对应的参考光流特征信息和预测光流特征信息之间的欧氏距离，作为每组视频帧对应的光流损失信息。
  12. 一种非临时性计算机可读存储介质，当所述存储介质中的指令由计算机设备的处理器执行时，使得所述计算机设备能够执行如权利要求1-6中任 一项所述的视频动作分类的方法。
PCT/CN2019/106250 2018-11-28 2019-09-17 视频动作分类的方法、装置、计算机设备和存储介质 Ceased WO2020108023A1 (zh)

Priority Applications (1)

Application Number Priority Date Filing Date Title
US17/148,106 US20210133457A1 (en) 2018-11-28 2021-01-13 Method, computer device, and storage medium for video action classification

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
CN201811437221.XA CN109376696B (zh) 2018-11-28 2018-11-28 视频动作分类的方法、装置、计算机设备和存储介质
CN201811437221.X 2018-11-28

Related Child Applications (1)

Application Number Title Priority Date Filing Date
US17/148,106 Continuation US20210133457A1 (en) 2018-11-28 2021-01-13 Method, computer device, and storage medium for video action classification

Publications (1)

Publication Number Publication Date
WO2020108023A1 true WO2020108023A1 (zh) 2020-06-04

Family

ID=65383112

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2019/106250 Ceased WO2020108023A1 (zh) 2018-11-28 2019-09-17 视频动作分类的方法、装置、计算机设备和存储介质

Country Status (3)

Country Link
US (1) US20210133457A1 (zh)
CN (1) CN109376696B (zh)
WO (1) WO2020108023A1 (zh)

Cited By (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN112966584A (zh) * 2021-02-26 2021-06-15 中国科学院上海微系统与信息技术研究所 一种运动感知模型的训练方法、装置、电子设备及存储介质

Families Citing this family (12)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN109376696B (zh) * 2018-11-28 2020-10-23 北京达佳互联信息技术有限公司 视频动作分类的方法、装置、计算机设备和存储介质
CN109992679A (zh) * 2019-03-21 2019-07-09 腾讯科技（深圳）有限公司 一种多媒体数据的分类方法及装置
CN110766651B (zh) * 2019-09-05 2022-07-12 无锡祥生医疗科技股份有限公司 超声设备
CN111241985B (zh) * 2020-01-08 2022-09-09 腾讯科技（深圳）有限公司 一种视频内容识别方法、装置、存储介质、以及电子设备
CN112784704A (zh) * 2021-01-04 2021-05-11 上海海事大学 一种面向小样本视频动作分类的方法
CN115018887B (zh) * 2021-03-04 2025-03-11 北京字节跳动网络技术有限公司 应用程序性能测试方法、装置、设备以及介质
CN114245206B (zh) * 2022-02-23 2022-07-15 阿里巴巴达摩院(杭州)科技有限公司 视频处理方法及装置
CN114782860B (zh) * 2022-03-29 2024-11-08 东北大学 一种监控视频中暴力行为检测系统及方法
CN115130539B (zh) * 2022-04-21 2024-11-08 腾讯科技（深圳）有限公司 分类模型训练、数据分类方法、装置和计算机设备
CN116486472B (zh) * 2022-12-29 2025-11-18 超级视线科技有限公司 监控视频的群体行为识别方法以及系统
CN116343134A (zh) * 2023-05-30 2023-06-27 山西双驱电子科技有限公司 驾考车辆信号传输系统及方法
US20250168367A1 (en) * 2023-11-16 2025-05-22 Disney Enterprises, Inc. Tunable Hybrid Neural Video Representations

Citations (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20120219186A1 (en) * 2011-02-28 2012-08-30 Jinjun Wang Continuous Linear Dynamic Systems
CN104966104A (zh) * 2015-06-30 2015-10-07 孙建德 一种基于三维卷积神经网络的视频分类方法
CN106599789A (zh) * 2016-07-29 2017-04-26 北京市商汤科技开发有限公司 视频类别识别方法和装置、数据处理装置和电子设备
CN107169415A (zh) * 2017-04-13 2017-09-15 西安电子科技大学 基于卷积神经网络特征编码的人体动作识别方法
CN109376696A (zh) * 2018-11-28 2019-02-22 北京达佳互联信息技术有限公司 视频动作分类的方法、装置、计算机设备和存储介质

Family Cites Families (11)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US7535463B2 (en) * 2005-06-15 2009-05-19 Microsoft Corporation Optical flow-based manipulation of graphical objects
CN105389567B (zh) * 2015-11-16 2019-01-25 上海交通大学 基于稠密光流直方图的群体异常检测方法
CN105956517B (zh) * 2016-04-20 2019-08-02 广东顺德中山大学卡内基梅隆大学国际联合研究院 一种基于密集轨迹的动作识别方法
US10402697B2 (en) * 2016-08-01 2019-09-03 Nvidia Corporation Fusing multilayer and multimodal deep neural networks for video classification
CN106599907B (zh) * 2016-11-29 2019-11-29 北京航空航天大学 多特征融合的动态场景分类方法与装置
CN106980826A (zh) * 2017-03-16 2017-07-25 天津大学 一种基于神经网络的动作识别方法
CN110622169A (zh) * 2017-05-15 2019-12-27 渊慧科技有限公司 用于视频中的动作识别的神经网络系统
CN108229338B (zh) * 2017-12-14 2021-12-21 华南理工大学 一种基于深度卷积特征的视频行为识别方法
CN108648746B (zh) * 2018-05-15 2020-11-20 南京航空航天大学 一种基于多模态特征融合的开放域视频自然语言描述生成方法
US11521044B2 (en) * 2018-05-17 2022-12-06 International Business Machines Corporation Action detection by exploiting motion in receptive fields
US11016495B2 (en) * 2018-11-05 2021-05-25 GM Global Technology Operations LLC Method and system for end-to-end learning of control commands for autonomous vehicle

Patent Citations (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20120219186A1 (en) * 2011-02-28 2012-08-30 Jinjun Wang Continuous Linear Dynamic Systems
CN104966104A (zh) * 2015-06-30 2015-10-07 孙建德 一种基于三维卷积神经网络的视频分类方法
CN106599789A (zh) * 2016-07-29 2017-04-26 北京市商汤科技开发有限公司 视频类别识别方法和装置、数据处理装置和电子设备
CN107169415A (zh) * 2017-04-13 2017-09-15 西安电子科技大学 基于卷积神经网络特征编码的人体动作识别方法
CN109376696A (zh) * 2018-11-28 2019-02-22 北京达佳互联信息技术有限公司 视频动作分类的方法、装置、计算机设备和存储介质

Cited By (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN112966584A (zh) * 2021-02-26 2021-06-15 中国科学院上海微系统与信息技术研究所 一种运动感知模型的训练方法、装置、电子设备及存储介质
CN112966584B (zh) * 2021-02-26 2024-04-19 中国科学院上海微系统与信息技术研究所 一种运动感知模型的训练方法、装置、电子设备及存储介质

Also Published As

Publication number Publication date
CN109376696B (zh) 2020-10-23
CN109376696A (zh) 2019-02-22
US20210133457A1 (en) 2021-05-06

Similar Documents

Publication Publication Date Title
WO2020108023A1 (zh) 视频动作分类的方法、装置、计算机设备和存储介质
CN108234870B (zh) 图像处理方法、装置、终端及存储介质
CN110249622B (zh) 实时的语义感知的相机曝光控制
TWI721603B (zh) 資料處理方法和資料處理裝置、電子設備和電腦可讀儲存媒體
WO2020125372A1 (zh) 混合声音信号的分离方法、装置、电子设备和可读介质
US10070050B2 (en) Device, system and method for cognitive image capture
CN105981368B (zh) 在成像装置中的照片构图和位置引导
KR20250157522A (ko) 자동화된 이미지 생성을 위한 프롬프트 수정
WO2020134556A1 (zh) 图像风格迁移方法、装置、电子设备及存储介质
CN106446782A (zh) 图像识别方法及装置
CN108174096A (zh) 拍摄参数设置的方法、装置、终端及存储介质
CN113259583B (zh) 一种图像处理方法、装置、终端及存储介质
CN107871001B (zh) 音频播放方法、装置、存储介质及电子设备
TWI735112B (zh) 圖像生成方法、電子設備和儲存介質
CN109819288A (zh) 广告投放视频的确定方法、装置、电子设备及存储介质
CN111242176A (zh) 计算机视觉任务的处理方法、装置及电子系统
US20150189166A1 (en) Method, device and system for improving the quality of photographs
CN115767256A (zh) 拍摄方法、装置、电子设备和存储介质
CN118401966A (zh) 动作识别机器学习模型的协同训练
CN116010646A (zh) 视频推荐方法、装置、电子设备及存储介质
CN108668061A (zh) 智能照相机
US9992407B2 (en) Image context based camera configuration
CN109460458B (zh) 查询改写意图的预测方法及装置
WO2025092911A1 (zh) 视频特征提取方法、视频生成方法、装置、介质及设备
CN107526438B (zh) 根据举手动作进行跟踪录播的方法、装置及存储装置

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: 19889581

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: 19889581

Country of ref document: EP

Kind code of ref document: A1