Versions Compared

Old Version 28

changes.mady.by.user Darya Andryieuskaya

Saved on

compared with

New Version 29

changes.mady.by.user Bahdan Tsiulkou

Saved on

Key

  • This line was added.
  • This line was removed.
  • Formatting was changed.
Tip

Hardware requirements for the Face detector VA and its sub-detectors

Hardware requirements for neural analytics operation

Video stream and scene requirements for the Face detector VA and its sub-detectors

Image requirements for the Face detector VA and its sub-detectors

Optimizing the operation of the Face detector VA on GPU in Windows OS

Checking in lists of faces

Examples of macros used when working with lists of faces

Note
titleAttention!

To start and correctly operate the Face detector VA on GPU, you must create cache beforehand (see Optimizing the operation of the Face detector VA on GPU in Windows OS).

To configure the Face detector VA, do the following:

  1. Go to the Detectors tab.

  2. Below the required camera, click Create… → Category: Face → Face detector VA.

By default, the detector is enabled and set to detect faces.

If necessary, you can change the detector parameters. The list of parameters is given in the table:

ParameterValueDescription
Object features
Check in listsYesThe
Check in lists
parameter is disabled by default. If you want to use this detector to check in lists of faces, select the Yes value (
see 
see Checking in lists of faces)
No
Record objects trackingYesThe metadata of the video stream is recorded to the database by default. To disable the parameter, select the No value
No
Video streamMain stream

If the camera supports multistreaming, select the stream for which detection is needed. For the correct operation of the Face detector

VA

TV, we recommend using a high-quality video stream

Second stream
Other
EnableYesThe detector is enabled by default. To disable the detector, select the No value
No
NameFace detector
VA
TVEnter the detector name or leave the default name
Decoder modeAuto

Select a processor for decoding video. When you select:

  • Auto: GPU takes priority (decoding with Nvidia NVDEC chips). If there is no appropriate GPU, the decoding
will use
  • uses the Intel Quick Sync Video technology. Otherwise, CPU resources
will be
  • are used for decoding;
  • CPU:
CPU Liveness detection
  •  CPU is used for decoding;
  • GPU: GPU is used for decoding (decoding with Nvidia NVDEC chips);
  • HuaweiNPU: HuaweiNPU is used for decoding
CPU
GPU
HuaweiNPU
TypeFace detector TV

Name of the detector type (non-editable field)

Use camera transform

Yes
The Liveness detection parameter
The parameter is disabled by default.
Liveness detection is used when there is a photo of a face instead of an alive person. If you want to save to the database the information about whether each captured face is a photo (Yes/No), select the Yes value
If you use a XingYun bispherical lens (see Configuring fisheye cameras), by default the detector receives the image of two spheres of 180° each to analyze. In this case, the recognition quality can deteriorate. To send the dewarped image to the detector, select the Yes value. This parameter is also valid for other transformations
No

Advanced settings

Note
titleAttention!

You must perform the advanced configuration of the detector only with the assistance of AxxonSoft technical experts.

Analyze face rotation angleYes

The parameter is disabled by default. If you want to detect the face rotation angle, select the Yes value. This parameter allows you to filter out results that have a rotation and tilt angle greater than the specified values in a search for a specific face (see Search for similar face)

Info
titleNote

The Analyze face rotation angle parameter affects the filtering of face detection events in the Event Board (see Configuring an Event Board, Working with Event Boards).

When you use the following filters:

  • Triggered detection—all detected faces are displayed regardless of the face angle settings, even if the Analyze face rotation angle parameter is enabled.
  • Triggered specified detection—only faces that are within allowable rotation angles specified in the settings of the detector are displayed.
No
Face detection algorithmALG1 (high speed, low accuracy)

Select the face detection algorithm:

  • ALG1 (high speed, low accuracy) is the fastest, reduced accuracy. Recognition speed depends on the number of faces in the frame. Optimal for many scenes.
  • ALG2 (medium speed, medium accuracy)—recognition speed depends on the background and the number of faces in the frame.
  • ALG3 (low speed, high accuracy) is the most accurate, reduced speed. Recognition speed depends on the image resolution
ALG2 (medium speed, medium accuracy)
ALG3 (low speed, high accuracy)
Face detection period (msec)250

Specify the time in milliseconds between face search operations in a video frame. The larger the value, the less load on the server, but some faces may not be recognized. The value must be in the range [1, 10 000]

Face rotation pitch ( ° )

45Specify the allowable angle of face pitch in degrees. You must select the required value empirically. The value must be in the range [0, 90]
Face rotation roll ( ° )45

Specify the allowable angle of face roll in degrees. You must select the required value empirically. The value must be in the range [0, 90]

Face rotation yaw from ( ° )

-45

Specify the minimum allowable angle of face rotation to the right or left. You must select the required value empirically. The value must be in the range [-90, 90]

Face rotation yaw to ( ° )

45

Specify the maximum allowable angle of face rotation to the right or left. You must select the required value empirically

NoLiveness threshold30

Specify the threshold value in percent at which a face is defined as liveless. The higher the value, the fewer captured faces are detected as a photo. At the same time, the quality of recognizing whether the captured face is a photo (Yes/No) is higher. You must select the required value empirically. The value must be in the range [1, 100]

TypeFace detector VA

Name of the detector type (non-editable field)

Use camera transformYesThe Use camera transform parameter is disabled by default. If you use a XingYun bispherical lens (see Configuring fisheye cameras), by default the detector receives the image of two spheres of 180° each to analyze. In this case, the recognition quality can deteriorate. To send the dewarped image to the detector, select the Yes value. This parameter is also valid for other transformationsNo

Advanced settings

Note
titleAttention!

Advanced configuration of the detector must be performed only with the assistance of AxxonSoft technical experts.

Analyze face rotation angleYes

The Analyze face rotation angle parameter is disabled by default. If you want to detect the face rotation angle, select the Yes value. This parameter allows you to filter out results that have a rotation and tilt angle greater than the specified values in a search for a specific face (see Search for similar face)

Info
titleNote

The Analyze face rotation angle parameter affects the filtering of face detection events in the Event Board (see Configuring an Event Board, Working with Event Boards).

When you use the following filters:

  • Triggered detection—all detected faces are displayed regardless of the face angle settings, even if the Analyze face rotation angle parameter is enabled.
  • Triggered specified detection—only faces that are within allowable rotation angles specified in the settings of the detector are displayed.
NoFace detection algorithmALG1 (high speed, low accuracy)

Select the face detection algorithm:

  • ALG1 (high speed, low accuracy) is the fastest, reduced accuracy. Recognition speed depends on the number of faces in the frame. Optimal for many scenes.
  • ALG2 (medium speed, medium accuracy)—recognition speed depends on the background and the number of faces in the frame.
  • ALG3 (low speed, high accuracy) is the most accurate, reduced speed. Recognition speed depends on the image resolution
ALG2 (medium speed, medium accuracy)ALG3 (low speed, high accuracy)Face detection period (msec)250Specify the time in milliseconds between face search operations in a frame. The larger the value, the less load on the Server, but some faces may not be recognized

. The value must be in the range [

1

-90,

10 000

90]

Face rotation pitch ( ° )

45

Specify the allowable angle of face pitch in degrees. You must select the required value empirically. The value must be in the range [0, 90]

Face rotation roll ( ° )45

Specify the allowable angle of face roll in degrees. You must select the required value empirically. The value must be in the range [0, 90]

Face rotation yaw from ( ° )

-45

Specify the minimum allowable angle of face rotation to the right or left. You must select the required value empirically. The value must be in the range [-90, 90]

Face rotation yaw to ( ° )

45
False mask detections filteringYesThe parameter is enabled by default. The additional algorithm considers that the track is a face of a person with a mask. If the algorithm considers that the track is a face with a mask with a lower probability, this track is disregarded (see the Minimum filtering threshold for face mask detection parameter). If you don’t need to use the parameter, select the No value
No
Filter false alarmsYesIn some cases, the detector can mistake other objects for a face. The Filter false alarms parameter is enabled by default. If the parameter is enabled, false results of the face recognition appear in the detection feed (see Face recognition and search), but they are ignored when searching for faces in the archive. This parameter filters out objects that aren't faces at the stage of face vector construction and its recording to the metadata database. If you don’t need to use this parameter, select the No value
No
Frame size change1920

By default, during the analysis, the frame is compressed to the specified resolution (1920 pixels on the longer side, by default).

Specify the maximum allowable angle of face rotation to the right or left. You must select the required value empirically.

The value must be in the range [

-90

640,

90]

False mask detections filtering

Yes

The False mask detections filtering parameter is enabled by default. Additional algorithm considers that the track is a face of a person with a mask. If algorithm considers that the track is a face with a mask with a lower probability, this track is disregarded (see the Minimum filtering threshold for face mask detection parameter). If you don’t need to use the parameter, select the No value

NoFilter false alarmsYes

In some cases, the detector can mistake other objects for a face. The Filter false alarms parameter is enabled by default. If the parameter is enabled, false results of the face recognition appear in the detection feed (see Face recognition and search), but they are ignored when searching for faces in the archive. This parameter filters out objects that are not faces at the stage of face vector construction and its recording to the metadata database. If you don’t need to use this parameter, select the No value

NoFrame size change1920

The analyzed frames are scaled down to a specified resolution (1920 pixels on the longer side) by default. The value must be in the range [640, 10 000]. The following algorithm is used:

  • If the longer side of the source image exceeds the value specified in the Frame size change field, it is divided by two.
  • If the resulting resolution falls below the specified value, the algorithm stops and this resolution will be used further.
  • If the resulting resolution still exceeds the specified value, it is divided by two until it is less than the specified resolution.

10 000]. The following algorithm is used:

  • If the longer side of the source image exceeds the value specified in the Frame size change field, it is divided by two.
  • If the resulting resolution falls below the specified value, the algorithm stops, and this resolution is used further.
  • If the resulting resolution still exceeds the specified value, it is divided by two until it is less than the specified resolution
Info
titleNote

For example, the source image resolution is 2048*1536, and the specified value is set to 1000. In this case, the source resolution is halved two times (512*384), since after the first division, the number of pixels on the longer side exceeds the specified value (1024 > 1000). If detection is performed on a higher resolution stream and detection errors occur, we recommend reducing the compression.

Ignore repeated recognitionsYesThe parameter is disabled by default. If you want to ignore repeated recognitions of the same face, select the Yes value
No
Liveness threshold
30

Specify the minimum threshold value in percent at which a face is defined as liveless. The higher the value, the fewer captured faces are detected as a photo. At the same time, the quality of recognizing whether the captured face is a photo (Yes/No) is higher. You must select the required value empirically. The value must be in the range [1, 100]

Minimum face masking threshold70Specify the minimum threshold for recognizing a mask on the face. You must select the required value empirically. We recommend specifying the value of not less than 70. The value must be in the range [1, 100]
Minimum face quality for face mask detection30

Specify the minimum quality of the face image when recognizing masks (see Mask detector TV). You must select the required value empirically. We recommend specifying the value of not less than 30. The value must be in the range [1, 100]

Minimum face recognition quality50

Specify the minimum quality of the face image when recognizing faces without masks. You must select the required value empirically. We recommend specifying the value of not less than 50. The value must be in the range [1, 100]

Minimum filtering threshold50

If you use the Filter false alarms parameter, specify the minimum percentage at which the additional algorithm considers the track to be a person's face. If the algorithm considers the track to be a face with a lower probability, this track is disregarded

Info
titleNote
For example, the source image resolution is 2048*1536, and the specified value is set to1000. In this case, the source resolution will be halved two times (512*384), as after the first division, the number of pixels on the longer side exceeds the specified value (1024 > 1000). If detection is performed on a higher resolution stream and detection errors occur, we recommend reducing the compression.Ignore repeated recognitionsYes

The Ignore repeated recognitions parameter is disabled by default. If you want to ignore repeated recognitions of the same face, select the Yes value

NoMinimum face masking threshold70Specify the minimum threshold for recognizing a mask on the face

. You must select the required value empirically. We recommend specifying the value of not less than

70

50. The value must be in the range [1, 100]

Minimum
face quality
filtering threshold for face mask detection30
Specify the minimum quality of the face image when recognizing masks (see Mask detector VA)

If you use the False mask detections filtering parameter, specify the minimum percentage at which the additional algorithm considers the track to be a person's face with a mask. If the algorithm considers the track to be a face with a mask with a lower probability, this track will be disregarded. You must select the required value empirically. We recommend specifying the value of not less than 30. The value must be in the range [1, 100]

Minimum face recognition quality50

Period of ignoring repeated recognitions

2

To adjust this parameter, you must enable the Ignore repeated recognitions parameter. Specify the period in minutes during which new recognized faces are compared with the previous ones to detect similarities. The

Specify the minimum quality of the face image when recognizing faces without masks. You must select the required value empirically. We recommend specifying the value of not less than 50. The

value must be in the range [

1

2,

100

30]

Minimum filtering threshold50
Process color framesYesA black and white frame is processed by default. If you want the detector to use a color frame for processing, select the Yes value
No

Repeated recognitions similarity threshold

85

To adjust this parameter, you must enable the Ignore repeated recognitions parameter. Specify the similarity threshold of the face to the previously recognized faces in percent. If the similarity threshold is lower than the specified value, the face is recognized as a new one. The value must be in the

If you enable the Filter false alarms parameter, specify the minimum percentage at which the additional algorithm considers the track to be a person's face. If the algorithm considers the track to be a face with a lower probability, this track is disregarded. You must select the required value empirically. We recommend specifying the value of not less than 50. The value must be in the

range [1, 100]

Minimum filtering threshold for face mask detection30

If you enable the False mask detections filtering parameter, specify the minimum percentage at which the additional algorithm considers the track to be a person's face with a mask. If the algorithm considers the track to be a face with a mask with a lower probability, this track is disregarded. You must select the required value empirically. We recommend specifying the value of not less than 30. The value must be in the range [1, 100]

Period of ignoring repeated recognitions

2

To adjust this parameter, you must enable the Ignore repeated recognitions parameter. Specify the period in minutes during which new recognized faces are compared with the previous ones to detect similarities. The value must be in the range [2, 30]

Process color framesYes

A black and white frame is processed by default. If you want the detector to use a color frame for processing, select the Yes value

No

Repeated recognitions similarity threshold

85

To adjust this parameter, you must enable the Ignore repeated recognitions parameter. Specify the similarity threshold of the face to the previously recognized faces in percent. If the similarity threshold is lower than the specified value, the face is recognized as a new one. The value must be in the range [1, 100]

Send face imagesYes

The Send face images parameter is disabled by default. If you want to send face images to Axxon PSIM, select the Yes value

NoTrack loss time500

Specify the time in milliseconds after which the track of the captured face is considered to be lost. This parameter applies when a captured face moves in the frame and gets hidden behind an obstacle for some time. If this time is less than the specified value, the face is recognized as the same. The value must be in the range [1, 10 000]

Basic settingsAge and genderYes

The Age and gender parameter is disabled by default. If you need to save age and gender information for each captured face in the database, select the Yes value

Info
titleNote

The average error in age recognition is 5 years.

NoAlgorithm of liveness detection
Not selected

The Not selected value is selected by default. Liveless face is a photo of a face instead of a live person. If necessary, select the algorithm for detecting a liveless face:

  • Algorithm 1 (Photo I) is used for processors with small processing power in access control systems (ACS). A simplified, fast box model is used to speed up processing.
  • Algorithm 2 (Photo I, Textures I) is used for most other ACSs, including a combination of a fast box model and texture model.
    • Send face imagesYesThe parameter is disabled by default. If you want to send face images to Axxon PSIM, select the Yes value
    No
    Track lifespan (starting with Detector Pack 3.14)Yes

    By default, the parameter is disabled. If it is necessary to display track lifespan for an object in seconds, select the Yes value

    No
    Track loss time500

    Specify the time in milliseconds after which the track of the captured face is considered to be lost. This parameter applies when a captured face moves in the frame and hides behind an obstacle for some time. If this time is less than the specified value, the face is recognized as the same thing. The value must be in the range [1, 10 000]

    Basic settings
    Age and genderYes

    The parameter is disabled by default. If you need to save age and gender information for each captured face in the database, select the Yes value

    Info
    titleNote

    The average error in age recognition is 5 years.

    No
    Algorithm of liveness detectionNot selected
    Note
    titleAttention!

    The algorithm is case-sensitive to the face position in the frame as well as the optic distortion. The algorithm operates correctly only when placing a face in the center of the frame area, where distortion is minimal. The lower the distortion in the frame, the higher the accuracy of the algorithm operation.

    The Not selected value is selected by default. Liveless face is a photo of a face instead of a live person. You can use a photo, mask, or video as an image. If necessary, select the algorithm for detecting a liveless face:

    • Algorithm 1 (Photo I) is used for processors with small processing power in access control systems (ACS). A simplified, fast box model is used to speed up processing.
    • Algorithm 2 (Photo I, Textures I) is used for most other ACSs, including a combination of a fast box model and texture model.
    • Algorithm 3 (Photo II, Textures II) is
    Algorithm 3 (Photo II, Textures II) is
    • used for remote authorization scenarios. It uses a more detailed and slower box model with a slower texture model, which improves verification accuracy.
    • Algorithm 4 (Comprehensive) is a mode for fine-tuning for specific conditions. We recommend using this mode if you can adjust the parameters to a specific domain and select threshold values. This approach
    improves
    • increases accuracy, but the threshold can vary significantly depending on the conditions.

    You must configure the parameter empirically.

    Algorithm for determining a liveless face:

    The algorithm assigns to each face a Liveness score—a value in the range [0, 100]. The decision on the liveness of a face is made based on the value specified in the Liveness threshold parameter according to the following logic:

    • liveless if a Liveness score is less than or equal to the threshold value of the parameter.
    • live if a Liveness score exceeds the threshold value of the parameter.

    Example:

    1. The Liveness threshold parameter = 30.
    2. In the logs, the algorithm assigned a Liveness score = 78.
    3. Since 78 > 30, the face is determined as live
    Algorithm 1 (Photo I)
    Algorithm 2
    (Photo I, Textures I)
    (Photo I, Textures I)
    Algorithm 3 (Photo II, Textures II)
    Algorithm 4 (Comprehensive)
    Biometric dataYes

    The parameter is enabled by default. When you search for faces (see Face search), the search results display found faces that are similar with an attached photo or track, taking into account the specified minimal level of similarity (in percent). For the correct check-in lists of faces, you must enable this parameter.
    If you want to keep the biometric data, select the No value. In this case, when you search for faces     (see Face search) with

    Algorithm 3 (Photo II, Textures II)Algorithm 4 (Comprehensive)Biometric dataYesThe Biometric data parameter is disabled by default. When you search for faces (see Face search) by

    an attached photo or track

    ,

    as well as

    when you

    check-in lists of faces,

    there will be no results

    it doesn't give results

    No
    Detection mode
    		CPU

    Select a processor for the detector operation—CPU or Nvidia GPU (see Selecting Nvidia GPU when configuring detectors)

    Note
    titleAttention!

    It can take several minutes to launch the algorithm on an Nvidia GPU after you apply the settings.

     

    Nvidia GPU 0
    Nvidia GPU 1
    Nvidia GPU 2
    Nvidia GPU 3
    Huawei NPU
    Face attributes recognition

    If you want to keep biometric data, select the Yes value. In this case, when you search for faces (see Face search), the search results will contain faces that are similar to the attached photo or track based on a specified minimum similarity level (percentage). For the correct check in lists of faces, this parameter must also be enabled

    NoFace attributes recognition 
    βeta
    		Not selected

    The Not selected value is selected by default. If necessary, select the

    algorithm of the face

    βeta face attributes recognition

    βeta

    algorithm:

    • Algorithm 1 (Fast recognition) is used for fast recognition and assessment.
    • Algorithm 2 (Accurate recognition)
     is
    • is used for a more accurate recognition and assessment
    Algorithm 1 (Fast recognition)
    Algorithm 2 (Accurate recognition)
    Face mask detectionYesThe
    Face mask detection
    parameter is disabled by default. If you use mask detector, select
    the 
    the Yes value (see Mask detector
    VA
    TV)
    No

    Maximum face height

    		100

    Specify the maximum height of the captured faces as a percentage of the frame size. You must select the required value empirically. The value must be in the range [1, 100]

    Maximum face width100

    Specify the maximum width of the captured faces as a percentage of the frame size.

     You

    You must select the required value empirically. The value must be in the range [1, 100]

    Minimum face height5Specify the minimum height of the captured faces as a percentage of the frame size.
     You
    You must select the required value empirically. The value must be in the range [1, 100]

    Minimum face width

    		5

    Specify the minimum width of the captured faces as a percentage of the frame size.

     You

    You must select the required value empirically. The value must be in the range [1, 100]

    Minimum threshold of face authenticity90

    Specify the minimum level of face recognition accuracy for the creation of a track. You

    must select the required value empirically. We recommend specifying the value of not less than 90. The higher the value, the fewer faces are detected, while the recognition accuracy increases. The value must be in the range [1, 100]Detection modeCPU

    Select a processor for the detector operation—CPU or Nvidia GPU (see Selecting Nvidia GPU when configuring detectors).

    Note
    titleAttention!

    It can take several minutes to launch the algorithm on Nvidia GPU after you apply the settings.

    Nvidia GPU 0Nvidia GPU 1Nvidia GPU 2Nvidia GPU 3Huawei NPU

    must select the required value empirically. We recommend specifying the value of not less than 90. The higher the value, the fewer faces are detected, while the recognition accuracy increases. The value must be in the range [1, 100]

    In If necessary, in the preview window, set the rectangular area of the frame in which you want to detect faces. You can specify the area by moving the anchor points points  Image Modified.

    Image Modified

    Info
    titleNote
    • For convenience of configuration, you can "freeze" the frame. Click the Image Modified button. To cancel the action, click this button again.
    • The detection area is displayed by default. To hide it, click the Image Modified button. To cancel the action, click this button again.

    To save the parameters of the detector, click the Apply Image Modified button. To cancel the changes, click the Cancel Image Modified button.

    Configuration of the Face detector VA is complete. If necessary, you can create and configure sub-detectors on the basis of the Face detector VA based on metadata (see Metadata database):

    1. Line crossing—detector generates an event when a person moves across a line in the specified area of the frame and their face is detected.
    2. Entrance in area—detector generates an event when a person appears in the specified area of the frame and their face is detected.
    3. Loitering in area—detector generates an event when a person stays in the specified area of the frame for a long time and their face is detected.
    4. Mask detector VA—detector generates an event when it captures a face with or without a mask.