You are using an unsupported browser. Please update your browser to the latest version on or before July 31, 2020.
close

IVA AI - Technical Note - DL Object Tracker

IVA AI – Technical Note – DL Object Tracker

-----------------------------------

Affected Roles:  Administrator, Owner

Related Digital Watchdog VMS Apps:  DW Spectrum® IPVMS

Software Version:  DW Spectrum IPVMS v5.0 and newer

Last Edit:  August 2, 2022

-----------------------------------

DL Object Tracker

This article will outline the practical applications and recommended practices for using the DL Object Tracker for your consideration.

Supported/Affected Devices:

  • DW Blackjack® Ai Servers

About This Feature

Practical Applications

The DL Object Tracker is designed to detect, track, and classify objects in a scene using a pre-trained model.

Example applications include:

  • Security
    • Perimeter protection
    • Monitoring areas with restricted access
    • Remote site security
    • Sterile zone (no entry)
    • Object detection (i.e., bag)

  • Retail
    • Object tracking using surveillance cameras
    • Accurate foot traffic tally counting
    • Dwell times (loitering)
    • Queue monitoring

  • Vehicle Traffic
    • Accurate vehicle tally counting
    • Stopped vehicle
    • Unauthorized vehicle detection
    • Rule violations

Limitations

The DL Object Tracker has been trained to work efficiently in many environments. However, the distance is defined by the lens and focal point of the camera. The DL Object Tracker will not provide accurate tracking of objects that are less than 25 ppm (pixels per meter).

The DL Object Tracker has not been trained to operate accurately with:

  • Thermal camera views
  • Fisheye cameras
  • Small pixel-sized objects (objects in the distance)
  • Situations where speed estimation is required

General Use Guidelines

The following guidelines should be followed to provide optimal object tracking and classification.

Camera Position and Angle

  • The camera should be mounted at least 2.8 meters from the ground plane
  • Cameras should be mounted on a stable surface to reduce the effects of vibration or other environmental factors that may affect the stability of the camera image
  • The camera view angle (tilt) should be within 30° from the horizontal
  • For high tracking accuracy, position the camera so that objects will be present within the camera view for at least 2 seconds

Obstructions

Camera scenes should be clear of foliage and other environmental factors that may obscure the camera view and reduce object detection and tracking.

  • Avoid installing the camera in locations that include large objects that may obscure a clear line of sight
  • Avoid installing the camera in locations that include constantly moving foliage that can interfere with tracking
  • Beware installing the camera where surfaces may reflect light (both white and infrared) into the camera lens

Lighting Conditions

  • Camera positions should be installed while avoiding allowing direct sunlight and other bright light sources from shining directly into the camera lens
  • Avoid placing cameras in locations that may experience drastic light changes
  • Avoid placing cameras in locations where the camera lens may be exposed to indirect light sources
    • Examples include direct sunlight, headlights, reflective surfaces and white light illuminators. These may result in poor object detection and tracking, reducing the effectiveness of video analytics
  • Minimum recommended lux on target is 10. Some cameras provide a reading of the lighting value and third-party tools are available to provide a lux reading
  • Bad weather environments will impact video analytics and can reduce object tracking accuracy

Frequently Asked Questions

  • Does the DL Object Tracker improve during operation?
    • It does not learn while in use. The DL Object Tracker detects and classifies objects using a pre-trained model. This is created using large sets of reference data.
  • Is calibration still required?
    • Because the DL Object Tracker detects and classifies objects that are in the scene, no calibration is currently required. As such, there are no calibration settings available when the DL Object Tracker is selected.
  • What objects will be detected?
    • The DL Object Tracker will only track objects that it can classify. Currently, these objects are:
      • Person
      • Cyclist
      • Car
      • Van
      • Truck
      • Motorcycle
      • Bicycle
      • Bag

  • How do I define the camera view?
    • The DL Object Tracker is designed to be used with a variety of standard CCTV camera views, so defining the positions should not be required.
    • It is recommended that you can define the scenario that you would like to create an event for to decide the best combination of trackers, zones, and analytic rules.
    • Ideally, the center of the camera view will be the focal point where the tracker will be the most active.
  • What distance will the DL Object Tracker detect?
    • Detection range will vary based on a camera’s field of view, which is defined by the camera’s lens and focal point. Instead, the DL Object Tracker detection is based on the number of pixels that make up an object. Currently, the approximate pixels per meter values are:
      • Person = 25 ppm
      • Vehicle = 30 ppm

  • What image settings should I use on the camera?
    • The image quality is an essential component to achieving accurate detection and tracking. The optimal resolution is 640x480 or D1 (720x480). Defining a lower resolution will reduce detection and tracking accuracy. Defining a larger resolution (such as 1920x1080) may provide a small increase in detection and tracking but will significantly increase the resource usage and reduce the system’s channel capacity.
    • The optimal framerate is 15 fps. Providing a lower framerate reduces the detection and tracking accuracy while increasing the framerate will only result in an increase in resource usage without providing an increase in detection and tracking.
    • The image quality and bitrate are important to ensure a good quality image is provided for video analytics. Where possible, the image quality and bitrate should be set to a High or Maximum setting with the bitrate defined near or at the maximum available limit. This ensures that a good image can be presented for the tracker to detect and track accurately.

Setup

The DL Object Tracker is active on the channel as soon as it is selected. On initial use, the tracker optimizes the available GPU(s). During this initialization process, a message will display on the channel preview page. This diagnostic process can take up to 45 minutes to complete.

The tracker assesses every frame being delivered regardless of activity. This means that the GPU loading will be constant.

General

  • Check that the required cameras and hardware are installed and working properly. The DL Object Tracker requires a Nvidia graphics card and CUDA libraries.
  • The DL Object Tracker is controlled through licensing. Ensure that the correct license has been activated on the system.
  • Licenses are assigned to channels. Ensure that the correct license is assigned to each channel.
  • Ensure that you have selected the correct tracker type for your application
  • Ensure a good quality image is being delivered to the analytic engine
  • Any object that can be classified will be identified and tracked. Use Object Filters to control which objects will trigger an event.
  • 875
  • 21-Nov-2022
  • 334 Views