Versions Compared

Old Version 2

changes.mady.by.user Sofya Uyutova

Saved on

compared with

New Version Current

changes.mady.by.user Alina Luchkina

Saved on

Key

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

The following detection tools  Axxon One x64 detectors grouped by tabs are available for Axxon Next platform calculation:

...

selection.

The Base tab

Base

Name

Description
Features of calculation

Motion Detection (CPU

, 20fps)

NOT a smart video detection tool (Motion detection).

-

Motion Detection (key frames)

NOT a smart video detection tool (Motion detection) with key frames decoding enabled.

-

)

Base motion detector when using the СPU resources. Changing the frame rate in the settings of the detector (the Frames processed per second parameter) does not significantly affect the load

Motion Detection (GPU)

Base motion detector when using the GPU resources. In this case, the GPU decoder operation mode was used. Changing the frame rate in the settings of the detector (the Frames processed per second parameter) does not significantly affect the load.

The models and the number of GPUs are selected separately

Service Detection (CPU, key frames)

Service detectors with decoding by key frames when using the СPU resources:

  • Quality degradation.
  • Blurred Image Detection.
  • Compression Artifacts Detection.
  • Image Noise Detection.
  • Scene change
Service detector (key frames)

Axxon Next detection tools:

  • Loss of quality;
  • Blurred Image Detection;
  • Compression Artifacts Detection;
  • Image Noise Detection;
    • Position change
    • .

    The platform is calculated for one service

    detection tool

    detector (any of the listed)

    Motion Detection (GPU, 20fps)

    .

    The results are given for decoding by key frames if the GOP=25 (every 25th frame is the key frame). The detector is applicable only for H.264, H.265 codecs

    Detection embedded in camera (CPU)Embedded detectors (built-in analytics) in camera when using the СPU resources

    The Tracker tab

    Name

    		Description
    Tracker VMDA (CPU)

    Scene analytics detectors (VMDA) based on object tracker when using the СPU resources.

    The results are given for the object tracker with one active Motion In Area sub-detector

    AI tracker with a neural filter (CPU)

    Scene analytics detectors (VMDA) based on object tracker using a neural filter and CPU resources. 

    The results are given for the object tracker with a neural filter and with one active Motion In Area sub-detector

    AI tracker with a
    Tracker
    Tracker VMDA

    Object trajectories detection tool.

    -AI tracker with 
    neural filter (GPU)
    Object trajectories detection tool

    Scene analytics detectors (VMDA) based on object tracker using a neural

    network 

    filter and

    using the

    GPU resources

    For each track, one image per second is sent to neural network for classification.

    • The NVIDIA GeForce GT 730 video card is capable of processing about 70* classifications** per second.
    • The NVIDIA GeForce GTX 1070 video card is capable of processing about 220*** classifications per second.
    • The NVIDIA Tesla P40 video card is capable of processing about 122**** classifications per second.

    • The Intel Neural Compute Stick 1 (movidius I) is capable of processing about 58***** classifications per second.

    • The Intel Neural Compute Stick 2 (movidius II) is capable of processing about 200***** classifications per second.

    Several video cards can be in use in one system.

    For example, if you need to track 9 persons per second on 10 cameras, GeForce GTX 1070 or similar video card is suitable.

    Up to two Intel Neural Compute Stick can be in use in one system.

    . In this case, the CPU decoder operation mode was used.

    The results are given for the object tracker with a neural filter with one active Motion In Area sub-detector.

    The models and the number of GPUs are selected separately

    Neurotracker (CPU, 6 FPS)

    Scene analytics detectors based on neurotracker using CPU resources and resource-intensive neural networks to detect people or vehicles.

    You can select the type of recognition object for the detector: Person, Person (top-down view), Vehicle.

    Relative accuracy: medium. Relative resource intensity: low.

    These neural networks are embedded in the product and can be trained on demand to detect different objects. The frame rate specified during the Neurotracker object configuration (the Frames processed per second parameter) is indicated in brackets. This is the number of FPS processed by the module; the frame rate of the incoming video stream is usually higher.

    The results are given for neurotracker with one active Motion In Area sub-detector

    Neurotracker (GPU, 6 FPS)

    Scene analytics detectors based on neurotracker using GPU resources and resource-intensive neural networks to detect people or vehicles.

    The GPU decoder operation mode was used.

    You can select the type of recognition object for the detector: Person, Person (top-down view), Vehicle.

    Relative accuracy: medium. Relative resource intensity: low.

    These neural networks are embedded in the product and can be trained on demand to detect different objects. The frame rate specified during the Neurotracker object configuration (the Frames processed per second parameter) is indicated in brackets. This is the number of FPS processed by the module; the frame rate of the incoming video stream is usually higher.

    The results are given for neurotracker with one active Motion In Area sub-detector

    Neurotracker (CPU, 6 FPS)—Person and Vehicle

    Scene analytics detectors based on neurotracker using CPU resources and high-precision neural network to detect people and (or) vehicles.

    You can select the type of recognition object and accuracy for the detector:

    • Nano: relative accuracy—moderately high, relative resource intensity—medium.
    • Medium: relative accuracy—high, relative resource intensity—high.

    These neural networks are embedded in the product and can be trained on demand to detect different objects. The frame rate specified during the Neurotracker object configuration (the Frames processed per second parameter) is indicated in brackets. This is the number of FPS processed by the module; the frame rate of the incoming video stream is usually higher.

    The results are given for neurotracker with one active Line Crossing sub-detector

    Neurotracker (GPU, 6 FPS)—Person and Vehicle

    Scene analytics detectors based on neurotracker using GPU resources and high-precision neural network to detect people and (or) vehicles. In this case, the GPU decoder operation mode was used.

    You can select the type of recognition object and accuracy for the detector:

    • Nano: relative accuracy—moderately high, relative resource intensity—medium.
    • Medium: relative accuracy—high, relative resource intensity—high.
    • Large: relative accuracy—very high, relative resource intensity—very high.

    These neural networks are embedded in the product and can be trained on demand to detect different objects. The frame rate specified during the Neurotracker object configuration (the Frames processed per second parameter) is indicated in brackets. This is the number of FPS processed by the module

    Neural tracker (CPU, 6fps)Scene Analytics tool based on neural trackingThe frame rate shown in parentheses is specified when configuring the Neurotracker module (with the Frame rate parameter). This is the number of frames per second processed by the module******

    ; the frame rate of the incoming video stream is usually higher.

    The results are given for neurotracker with one active Line Crossing sub-detector

    Neural
    tracker
    counter (
    VPU
    GPU,
    6fps
    1FPS)Scene
    Analytics tool
    analytics detector based on
    neural tracking
    Neural counter when using the
    Vision processing unit (VPU) resources1 x Mustang-V100-MX8 (Intel HDDL) card processes up to 60****** video channels regardless of video resolution
    GPU resources. The GPU decoder operation mode was used.
    The results are given for the Neural counter with one active Motion In Area sub-detector.
    The frame rate
    shown in parentheses is specified when configuring the Neurotracker module (with the Frame rate parameter)
    specified during the detector configuration (the Frames processed per second parameter) is indicated in brackets. This is the number of
    frames per second
    FPS processed by the module; the frame rate of the incoming video stream is usually higher. 
    You can select the type of recognition object for the detector: Person, Person (top-down view), Vehicle.

    LPR&Traffic tab

    Name

    		Description
    License plate recognition VT (CPU)

    License plate recognition VT detector when using the СPU resources

    License plate recognition RR (CPU)License plate recognition RR detector when using the СPU resources
    License plate recognition RR (GPU)License plate recognition RR detector when using the GPU resources
    Vehicle make and model recognition RR (CPU)

    Detector recognizes makes, models, type, color and running lights of RR vehicles when using СPU resources

    License Plate Recognition

    License plate recognition detection tool

    -FaceFace Search

    Face search detection tool

    -Fire&SmokeFire and smoke Detection (CPU)Fire and smoke detection based on neural networkIn order to enhance quality of operation and reduce CPU usage, it is recommended to use the detection tool with calculation on GPU.Fire and Smoke Detection (GPU)Fire and smoke detection based on neural network using the GPU resources.

    500 MB of video memory per detection type is required regardless of the number of channels. For example, for any number of smoke detection channels, 500 MB is required, and if the server has any number of both smoke and fire detector channels at the same time, a video card with at least 1 GB of memory should be in use.

    Several video cards can be in use in one system.

    If the Time between processed frames in seconds parameter is set to the default value (10 seconds), any NVIDIA graphics card compatible with the detection tool will be suitable (see the requirements in the Axxon Next User Guide).

    Behavior analyticsPose detection (CPU, 3fps)Pose detection based on neural network

    The number of specific pose detection tools created under the Pose detection object as well as average number of objects detected per certain time period do not affect the calculation results (except Close-standing people detection which contributes to the overall load).

    The platform is calculated for Delay between two measurements value of 333 ms (i.e. 3 fps which is different from the default value).

    Pose detection (VPU, 3fps)Pose detection based on neural network using the Vision processing unit (VPU) resources

    The number of specific pose detection tools created under the Pose detection object as well as average number of objects detected per certain time period do not affect the calculation results (except Close-standing people detection which contributes to the overall load).

    The platform is calculated for Delay between two measurements value of 333 ms (i.e. 3 fps which is different from the default value).

    If decoding is performed on CPU, 2x Intel Xeon Gold 6130T or 1х Intel core i7-8700 process up to 28 channels******

    Equipment detection (CPU, 1fps)

    Equipment detection (VPU, 1fps)

    Personal protection equipment (PPE) detection tool based on neural network-

    ...

    titleNote.
    Vehicle make and model recognition RR (GPU)Detector recognizes makes, models, type, color and running lights of RR vehicles when using СPU resources
    License plate, make and model recognition RR (CPU)License plate recognition RR with enabled Make and model recognition (MMR) detector when using СPU resources
    License plate, make and model recognition RR (GPU)License plate recognition RR with enabled Make and model recognition (MMR) detector when using GPU resources
    License plate recognition IV (CPU)License plate recognition IV detector when using СPU resources
    License plate recognition IV (GPU)License plate recognition IV detector when using GPU resources

    The Face tab

    Name

    		Description
    Facial recognition (CPU)

    Face detector when using СPU resources

    Facial recognition VA (GPU)

    Face detector when using GPU resources. The GPU decoder operation mode was used

    The Fire&Smoke tab

    Name

    		Description

    Fire detector (CPU, 0.1 FPS)

    Smoke detector (CPU, 0.1 FPS)

    Fire and smoke detectors based on neural network using CPU resources.

    The frame rate specified during the detector configuration (the Frames processed per second parameter) is indicated in brackets. This is the number of FPS processed by the module; the frame rate of the incoming video stream is usually higher

    Fire detector (GPU, 0.1 FPS)

    Smoke detector (GPU, 0.1 FPS)

    		Fire and smoke detectors based on neural network using GPU resources.
    The frame rate specified during the detector configuration (the Frames processed per second parameter) is indicated in brackets. This is the number of FPS processed by the module; the frame rate of the incoming video stream is usually higher

    The Behavior analytics tab

    Name

    		Description

    Visitors counter (CPU)

    		Visitors counter when using CPU resources. The results are given when frame rate in the settings of the detector (the Frames processed per second parameter) is 25
    Heat map (CPU)Heat map based on object tracker when using СPU resources
    Queue detector (CPU)Queue detector when using СPU resources
    Human pose detector (CPU, 3 FPS)

    Human pose detectors based on neural network using CPU resources.

    The frame rate specified during the detector configuration (the Frames processed per second parameter) is indicated in brackets. This is the number of FPS processed by the module; the frame rate of the incoming video stream is usually higher.

    The number of specific pose detectors created in the configuration for the Human pose detector parent object does not affect the calculation results (except for the Close-standing people detector)

    Human pose detector (GPU, 3 FPS)

    Human pose detectors based on neural network using resources of computer vision processor (GPU). In this case, the GPU decoder operation mode was used.

    The frame rate specified during the detector configuration (the Frames processed per second parameter) is indicated in brackets. This is the number of FPS processed by the module; the frame rate of the incoming video stream is usually higher.

    The number of specific human pose detectors created in the configuration for the Human pose detector parent object does not affect the calculation results (except for the Close-standing people detector).

    The models and the number of GPUs are selected separately .

    The results are given for standard neural network capable of detecting an object sized of at least 5% of the frame width/height. The results can differ for neural network capable of detecting smaller objects (since more resources are required)

    Equipment detector (CPU, 1 FPS)

    Personal protection equipment (PPE) detectors based on neural network using CPU resources. 

    The frame rate specified during the detector configuration (the Frames processed per second parameter) is indicated in brackets. This is the number of FPS processed by the module; the frame rate of the incoming video stream is usually higher.

    The results are given for a detector with five classification networks operating simultaneously when identifying equipment on each body part (head, torso, hands, legs, feet) in a gateway: at the entrance to the area in which the equipment is required, an employee lingers for 5-10 seconds during which the detector identifies the presence of the necessary equipment

    Equipment detector (GPU, 1 FPS)

    Personal protection equipment (PPE) detectors based on neural network using resources of computer vision processor (GPU). In this case, the GPU decoder operation mode was used.

    The frame rate specified during the detector configuration (the Frames processed per second parameter) is indicated in brackets. This is the number of FPS processed by the module; the frame rate of the incoming video stream is usually higher.

    The results are given for a detector with five classification networks operating simultaneously when identifying equipment on each body part (head, torso, hands, legs, feet) in a gateway: at the entrance to the area in which the equipment is required, an employee lingers for 5-10 seconds during which the detector identifies the presence of the necessary equipment. 

    The models and the number of GPUs are selected separately.
    If you use GPU, both segmenting neural network and classification neural networks are processed on it

    Meta-detector (GPU)Meta-detector based on neural network when using GPU resources. The results are given when frame rate in the settings of the detector (the Frames processed per second parameter) is 1. The frame rate specified during the detector configuration (the Frames processed per second parameter) is the number of FPS processed by the module; the frame rate of the incoming video stream is usually higher. 
    Crowd esimation VA (GPU)Crowd estimation detector based on neural network using GPU resources. The results are given when frame rate in the settings of the detector (the Frames processed per second parameter) is 0.017. The frame rate specified during the detector configuration (the Frames processed per second parameter) is the number of FPS processed by the module; the frame rate of the incoming video stream is usually higher. 

    * – The results are given for Core i5-3570 (3400 MHz) CPU and may vary depending on the CPU installed. For example, the Xeon Gold 6140 (2300 MHz) CPU allows 95 classifications** per second.

    ** – 1 classification per second is 1 object detected on video. For example, if average of 9 moving objects are simultaneously present on video from one camera, and there are 5 cameras in the system, use video card allowing 45 classifications per second.

    *** – The results are given for the Core i7-8700 (3200 MHz) CPU and may vary depending on the CPU installed.

    **** – 360 classifications per second were achieved in test utility on the 2x Intel Xeon Gold 6140 platform. In Axxon Next, up to 122 classifications per second were possible with 90% CPU utilization.

    ***** – The results are given for the Core i7-3770 (3400 MHz) CPU and may vary depending on the CPU installed.

    ...