Skip to content

PeopleApp

App Overview

The PeopleApp detect one or more physical objects from one categories within an image and return a box around each object, as well as a category label for each object.

Main Feature

  1. It detects people and notifies them if they have fallen or if they are wearing a hard hat or mask.

  2. Counts the number of people in the area of ​​interest in real time.

App Application Architecture

Architecture Overview

Det(People)─────Tracker(NvDCF)───┬─Cls1(Falldown for person Category)
                                 ├─Cls2(Helmet   for person Category)
                                 └─Cls3(Mask     for person Category)

PeopleDet   : {0:person}
FalldownCls : {0:falldown, 1:none}
HelmetCls   : {0:wear, 1:off, 2:unknown}
MaskCls     : {0:off, 1:wear, 2:unknown}

Model Infomation

  • PeopleDet

    1. Description

      Detecting person objects against diverse backgrounds.

    2. Version

      v1.6.3

    3. Dataset Properties

      Category name Number of Images Number of Instances
      person 55892 121616
      Total 55892 121616

    4. Performance

      Category name Recall 50:95 Precision 50:95
      person 99.2 94,9
      mean 99.2 94.9

  • FalldownCls

    1. Description

      For person objects, classify whether they have fallen or not.

    2. Version

      v2.3.0

    3. Dataset Properties

      Category name Number of Images Number of Instances
      falldown 15712 15712
      none 10828 10828
      Total 26540 26540

    4. Performance

      Category name Accuracy
      falldown 98.4
      none 99.2
      mean 98.9

  • HelmetCls

    1. Description

      Person objects are classified into three categories: wearing a helmet, taking off a helmet, or unknown.

    2. Version

      v2.1.0

    3. Dataset Properties

      Category name Number of Images Number of Instances
      wear 4576 4576
      off 5197 5197
      unknown 284 284
      Total 4576 4576

    4. Performance

      Category name Accuracy
      wear 98.0
      off 96.7
      unknown 37.9
      mean 95.6

  • MaskCls

    1. Description

      Person objects are classified into three categories: masked, unmasked, or unknown.

    2. Version

      v2.0.0

    3. Dataset Properties

      Category name Number of Images Number of Instances
      off 3723 3723
      wear 2624 2624
      unknown 3812 3812
      Total 10159 10159

    4. Performance

      Category name Accuracy
      off 90.9
      wear 89.0
      unknown 89.5
      mean 89.9

Input / Output

Input

Input Type : Image

Input Format : RGB

Input Resolution : 640 X 640 X 3 (W x H x C)

INPUT IMAGE

Output

Output Type : Lable, Boundary-box(Bbox), Confidence Scores, Track ID

Output Type Data Format
Lable str
Bbox list: [x: float, y:float, w:float , h:float ]
Confidence Scores float
Track ID int
Event str

OUTPUT IMAGE (in Viveex)

Software Integration

Runtime Engine:

  • Autocare Edge 1.7

Preferred System Spec:

  • OS : Ubuntu (debian)
  • GPU : NVIDIA RTX 3070, NVIDIA RTX 4090, etc(Nvidia).

Limitation & Warning

Small objects : The object size must be at least 5% of the total image size for detection.

Crowded objects : In the case of crowd objects, they can be recognized as a single object by the Iou algorithm.

Camera distortion : A skewed input may not fit the model's domain.

Dark or blurry objects : It may not pass the threshold and therefore go undetected.