To configure the neural tracker-based Scene Analytics detection tools, do the following:

1. Select the Neurotracker object. 
   
2. By default, metadata are recorded into the database. To disable metadata recording, select No from the Record object tracking (1) list.
3. If the camera supports multistreaming, select the stream for which detection is needed (2). 

4. To reduce the number of false positives from a fish-eye camera, you have to position it properly (3). For other devices, this parameter is not valid.
   

5. Select a processing resource for decoding video streams (4). When you select a GPU, a stand-alone graphics card takes priority (when decoding with NVIDIA NVDEC chips). If there is no appropriate GPU, the decoding will use the Intel Quick Sync Video technology. Otherwise, CPU resources will be used for decoding.
6. Set the Detection threshold for objects in percent (5). If the recognition probability falls below the specified value, the data will be ignored. The higher the value, the higher the accuracy, but some triggers may not be considered.

7. Set the frame rate value for the neural network to process per second (6). The higher the value, the more accurate tracking, but the load on the CPU is also higher.

   6 FPS or more is recommended. For fast moving objects (running individuals, vehicles), you should set the frame rate at 12 FPS or above (see Examples of configuring neural tracker for solving typical tasks).

   
   

8. Specify the Minimum number of detection triggers for the neural tracker to display the object's track (7). The higher the value, the more is the time interval between the object's detection and the display of its track on the screen. Low values may lead to false positives.

9. If you use a unique neural network, select the corresponding file (8).
   

   To train your neural network, contact AxxonSoft (see Data collection requirements for neural network training). A trained neural network for a particular scene allows you to detect only objects of a certain type (e.g. person, cyclist, motorcyclist, etc.). If the neural network file is not specified, the default file will be used, which is selected automatically depending on the selected object type (13) and the selected processor for the neural network operation (4). If you use a custom neural network, enter a path to the file. The selected object type is ignored when you use a custom neural network. To ensure the correct operation of the neural network on Linux OS, the corresponding file should be located in the /opt/AxxonSoft/DetectorPack/NeuroSDK directory.

   
   

10. You can use the neural filter to sort out certain tracks. For example, the neural tracker detects all freight trucks, and the neural filter sorts out only video recordings that contain trucks with cargo door open. To set up a neural filter, do the following:

    1. to use the neural filter, select Yes in the corresponding field (9).

    2. in the Neurofilter file field, select a neural network file (10).

    3. in the Neurofilter mode field, select a processor to be used for neural network work (11, see General Information on Configuring Detection).

11. Select the processor for the neural network—CPU, one of NVIDIA GPUs or one of Intel GPUs (12, see Hardware requirements for neural analytics operation, General Information on Configuring Detection).

    We recommend using the GPU. It may take several minutes to launch the algorithm on NVIDIA GPU after you apply the settings. You can use caching to speed up future launches (see Configuring the acceleration of GPU-based neuroanalytics). If Neural Tracker is running on GPU, object tracks may be lagging behind the objects. If this happens, set the camera buffer size to 1000 milliseconds (see The Video Camera Object).

    
    

12. In the Object type field (13), select the recognition object:

    1. Human.
    2. Human (top view).
    3. Vehicle.
    4. Human and Vehicle (Nano)—high speed, low accuracy, low processor load.
    5. Human and Vehicle (Medium)—medium speed, medium accuracy, medium processor load.

    6. Human and Vehicle (Large)—low speed, high accuracy, high processor load.

13. To enable the search for similar persons, in the Similitude search field (14), select Yes. It increases the CPU load.

    The Similitude search works only on tracks of people.

    
    

14. In the Time of processing similitude track (sec) field (15), set the time in the range [0; 3600] required for the algorithm to process the track to search for similar persons.
15. If you don't need to detect moving objects, select Yes in the Hide moving objects field (16). An object is treated as static if it does not change its position more than 10% of its width or height during its track lifetime.
    

16. If you don't need to detect static objects, select Yes in the Hide stationary objects field (17). This parameter lowers the number false positives when detecting moving objects. An object is considered stationary if it has not moved more than 10% of its width or height during the whole time of its track existence.

    If a stationary object starts moving, the detection tool will trigger, and the object will no longer be considered stationary.

    
    

17. If necessary, enable the Model quantization option (18). It allows you to reduce the consumption of the GPU processing power.

    AxxonSoft conducted a study in which a neural network model was trained to identify the characteristics of the detected object with quantization. The following results of the study were obtained: model quantization can lead to both an increase in the percentage of recognition and a decrease. This is due to the generalization of the mathematical model. The difference in detection ranges within ±1.5%, and the difference in object identification ranges within ±2%. Model quantization is only applicable to NVIDIA GPUs. The first launch of a detection tool with quantization enabled may take longer than a standard launch. If GPU caching is used, next time a detection tool with quantization will run without delay.

    
    

18. If necessary, select the class of the object to be detected (19).

19. In the Track retention time field, set a time interval in seconds after which the object track is considered lost (20). This helps if objects in scene temporarily overlap each other. For example, a larger vehicle may completely block the smaller one from view. 

20. By default, the entire FOV is a detection area. If you need to narrow down the area to be analyzed, you can set one or several detection areas in the preview window.

    The procedure of setting areas is identical to the primary tracker's (see Setting General Zones for Scene analytics detection tools). The only difference is that the neural tracker areas are processed while the primary tracker areas are ignored.

    
    

21. Click the Apply button.

22. The next step is to create and configure the necessary detection tools on the basis of neural tracker. The configuration procedure is the same as for the primary tracker (see Setting up Tracker-based Scene Analytics detection tools).

    To trigger a Motion in Area detection tool under a neural network tracker, an object should be displaced by at least 25% of its width or height in FOV. The abandoned objects detection tool works only with the primary tracker.