A Real-Time Intelligent System for Uniform Dress Classification

Project Title

A Real-Time Intelligent System for Uniform Dress Classification

Project Area of Specialization Artificial IntelligenceProject Summary

Security is a major issue all over the world which
requires object recognition based on real-time parameters. Detection and recognition of an object from a real-time video
irrespective of their geometry, scale, and the view is pretty
difficult. It becomes much more critical and essential when
dealing with sophisticated areas such as schools, hospitals etc..In this work, we introduced and developed an Intelligent DressUniform Identification System (IDUIS). The IDUIS performs object classification based on the uniform dress using artificial intelligence and high-performance computing system in a real environment. The IDUIS targets scale invariant features of the object and apply different machine learning approaches for classifications of the dress.

CONCLUSION
In this work, we developed an Intelligent Dress Uniform
Identification System (IDUIS). The IDUIS performs uniform
dress classification based scale-invariant features and artificial intelligence techniques. The system uses a high-performance computing machine having CPU and GPU cores. While testing the IDUS in a real-time environment results confirms that the system identifies dress uniform with an accuracy of 97.7%.

RESULT

	svm	knn	random forest
cpu time frame	534,33	201.22	110.750
cpu time frame	134.3	15.11	11.3
accuracy	72.67%	97.67%	52.2%

 cpu time framecpu time frameaccuracyProject Objectives

Uniform Dress Classification is a technique that is used to identify the dress cloths having a specific pattern and order.

Different departments give it high importance by the department following rules and regulation such as Army, School, etc.

Used for Security, Safety, Privacy, Identification, etc

[OBJECTIVE

•A high-performance system for generalizing templates (Dress Uniform) based matching.

•The system operates in real-time and uses an Artificial Intelligent Algorithm.]

Project Implementation Method

Object recognition is a field that performs image processing
using a computer machine to recognize and understands images
based on different shapes and features. Object recognition
techniques are becoming popular and are useful for intelligent
computer vision. It allows the computer system to understand
and identify specific objects and shapes in a real-time environment.
Generally object recognition is done using small part of
the image having certain features called template. Template
matching is used in object recognition when the standard
deviation of the template image compared to the source image
is small enough. A dense illustration of objects template is not
very useful because it needs allot of processing performance
and is cumbersome to identify for compelx shapes. Certain
features are extracted from the templates in order to describe
their shape, to compare it to the features of template objects,
or to partition objects into classes of different shapes.
Real-time image recognition by using object recognition can
be used for medical, security, surveillance, defense, agriculture
etc.. It can be used for the high-security purpose as it can
be used to identify different objects based on specific features.
The objects recognition algorithms employees feature-based
techniques and target machine learning and pattern
recognition. However, behavioral or quality characteristics are
sometimes discarded while extracting certain features. The
major difficulty in real-time object recognition is the potency
of similarity matches with respect to noise. Many frequently
used object recognition algorithms experience from a lack of
strength. Small dissimilarity in the data can drive them towards
uninformative values. Therefore a scale and shape invariant
feature based object recognition is required that perform object
recognition in a real environment.
In this work, we have presented an Intelligent Dress Uniform
Identification System (IDUIS). The IDUIS applies to
scale-invariant features and machine learning techniques for
classification and uses high performance computing machine
for processing. The system is tested using uniform dress identification
in a variant environment. The results show that the
the system identifies the templates with high speed and accuracy

•Manually operated

•Domain-specific

•Application specific

•Difficult to program

Not scalable to HPC

GP-GPU Cluster

 9 TeraFLOPS

CPU Octa Core

GPU GTX 1080 2560 Cuda cores

l32 Giga Byte Main Memory

512 Giga Byte SSD

Benefits of the Project

Security (we used for security reason)

Safety(also for safety person should be clearly identified)

Privacy

Identification, etc

Technical Details of Final Deliverable

System Architecture

The System Architecture uses the heterogeneous multi-core
processing system having an 8-core Intel processor and a GPU GTX1080 Accelerator of 2560 Cuda-cores. The processing the system has the capacity to perform of 8.9 trillion Floating Point Operating per Second (FLOPS) The system architecture used a Linux operating system that performs resource management.
In order to program the IDUIS, the system uses python
programming and machine learning libraries

Algorithm

The IDUIS algorithm block diagram is shown in Figure 1.
The algorithm takes real-time video from the camera source.
As we are dealing with real-time, therefore, selecting specific
signature signals is very important that can solve the problem
in a real-time environment. The algorithm performs filtering

and the extracts invariant features.
The algorithm uses SIFT/SURF features extractor that extracts the invariant features from the image shown in Figure 3.
The KNN, Random Forest, SVM, and CNN machine learning
solvers are used to identify the dress uniform templates.
SIFT and SURF techniques are used to extract invariant
features from an image. SIFT/SURF features detectors extract
the minima and maxima in scale space of these difference of-
Gaussian images. At each of these minima and maxima, a
a comprehensive model is fit to determine location, scale, and contrast, during which some features are discarded based on measures of their vulnerability. Once a stable feature has been detected, its dominant gradient orientation is obtained, and a key-point descriptor vector is formed from a grid of gradient histograms constructed from the gradients in the neighborhood of the feature. Key-point matching between images is performed
using a nearest-neighbor indexing method, followed by
a Hough transform that finds key-points that agree on potential object poses, and finally a solution for affine parameters,
which determines the specific location and orientation of each
recognized object.
The scale invariant features are then given to K Nearest
Neighbors (KNN), Random-Forest, Support Vector Machine
(SVM) and Convolution Neural Network (CNN) machine
learning solvers to train and classy the features.
The KNN machine learning solver that stores all available
cases and classifies new cases based on a similarity measure.
 

RESULTS AND COMPARISON

In order to validate our algorithm we executed real-time
uniform dress identification. To train the IDUIS we used 100
templates each for shirts and pants. Later IDUIS is tested on
a real-time environment.
The Real-time Performance is measured by applying four
different classifiers (SVM, KNN, Random-Forest and CNN)
on subject movements and found that the results are completely
different because of their prediction approach and
working mechanism.

Final Deliverable of the Project HW/SW integrated systemType of Industry Security Technologies Artificial Intelligence(AI)Sustainable Development Goals Peace and Justice Strong InstitutionsRequired Resources

Elapsed time in (days or weeks or month or quarter) since start of the project	Milestone	Deliverable
Month 1	on working	yes