Implementation of an artificial vision system to track objects using algorithms to extract global and local characteristics.

1. Abstract

Object recognition is the task of automatically finding and identifying objects in an image. Humans detect and recognize objects in images with extreme ease, even if the objects undergo variations in shape, size, location, color, texture, brightness or are partially obstructed. However this is a difficult task for computer vision. Therefore, various methods have been developed that attempt to achieve perfection in the human eye. In this project, the study of two image identification techniques (hu and Surf moments) was carried out, to determine what are the advantages that each one presents and the steps for its implementation.

2. Objective 

Reinforce the knowledge acquired in class about working with the stages of image preprocessing and extraction of local and global characteristics in order to carry out object tracking tasks.


3. Introduction

Today the use of technology is vital. Therefore, significant advances are made daily in the application of technologies in various sciences. One of the areas in which it is working is the creation of fast, expert and autonomous computers.

One of the most ambitious goals is to give computers the ability to relate to their environment in the same way that humans do, that is, through the senses. Providing sensory capacity to a computer is a complicated task. Accessory elements to the microprocessor are needed such as sensors, signal or noise cards, etc. In this way, there is a special interest in one of the five senses: the ability to see.

• Artificial Vision

Artificial vision refers to the technology (usually, visual implants) that allows blind people to see. The main aim of visual implants is to relay pictures to the brain using either cameras or photoreceptor arrays. There are different types of implants used to stimulate vision: retinal, cortical and biohybrid. None of the currently available technologies restores full vision, but they are often able to improve one’s ability to recognize shapes and movements.

The retina is a thin layer of neural tissue that lines the back wall inside the eye. Some of these cells act to receive light, while others interpret the information and send messages to the brain through the optic nerve. This is part of the process that enables you to see. In damaged or dysfunctional retinas, the photoreceptors stop working, causing blindness. By some estimates, there are more than 10 million people worldwide affected by retinal diseases that lead to loss of vision.

Until now, those who lost their vision to retinal disease would have had little hope of regaining it. But technological breakthroughs could soon give back the gift of sight. Several groups of scientists have already developed silicon microchips that can create artificial vision. In this article, we will examine how your retinas work and why blindness caused by retinal disease no longer means the loss of vision (Bonsor, 2000).

When addressing an object recognition problem, a basic five-step process is commonly followed:

a. Image acquisition
Capture the real world scene through sensors and digitize it for processing, storage and transmission.

b. Preprocessing
Applying contrast enhancement techniques, noise reduction, feature enhancement, etc., so that the image is suitable for the following steps.

c. Segmentation
Isolate the objects of interest in the image.

d. Feature extraction:
Numerically describe the nature of the segmented objects such as their shape, color, texture, etc.

e. Classification
Assign a class or category to each image object based on its features.

There are several techniques that use algorithms to identify images, in this project we used:

• Global characteritics

Moments Hu 

The characteristic shape of an object can be quantified by moments, which describe how the pixels of an object are distributed over the image plane. Moments must be invariant (ie, similar values for objects of the same type) to the geometric transformations (translation, rotation and scaling) that the objects can undergo and at the same time must be discriminant (ie, different values for objects of different types). These features are desirable for easier recognition of objects.

The moments of Hu are a set of seven invariant descriptors that quantify the shape of an object.

• Local characteristics

SURF  (Speeded Up Robust Feature)

SURF is a robust local feature detector, first introduced by Herbert Bay in 2006, that can be used in computer vision tasks such as object recognition or 3D reconstruction. It is largely inspired by the SIFT descriptor. The standard version of SURF is much faster than SIFT and its authors claim that it is more robust against transformations in different images than SIFT. SURF is based on the sums of Haar's 2D wavelet model responses and makes efficient use of integral images.

An integer approximation is used for the Hessaino detector determinant (Hessian blob detector), which can be calculated extremely quickly with an integral image since there are three integer operations. For features, use the sum of the Haar wavelet response around the point of interest. Again, these can be calculated with the help of the integral image.

4. Development and results

All the methods where implemented in OPEN CV:

• Global characteristics (Hu Moments)

HSV value acquisition for each figure, three figures with three diferents colors were analyce:

A tree_yellow
A house_red
A cloud_blue

Yellow Tree HSV

Blue Cloud HSV

Red House HSV

Detection of the images:

• Local characteristics (SURF)

For this method, three images (symbols) were selected:

National coat of arms of Ecuador

Real Madrid shield

Burger King Logo

5. Conclusion

New technologies are increasingly incorporated into our daily lives. Applications such as biometric fingerprint control or digital cameras to recognize faces are now a reality thanks to the advances in Computer Vision, one of the branches of Artificial Intelligence that has experienced the greatest growth in recent years.

This techniques are very important in the new reality that we are living. Thus is really important that people do research and get involved in the creation of new technologies, building on existing methods such as SURF or Hu moments. Which are really easy methods to implement.


6. Bibliography

Bonsor, K. (2000, octubre 16). How Artificial Vision Will Work. HowStuffWorks. https://science.howstuffworks.com/innovation/everyday-innovations/artificial-vision.htm

La visión por computador: Una disciplina en auge. (2012, abril 19). Informatica ++. http://informatica.blogs.uoc.edu/2012/04/19/la-vision-por-computador-una-disciplina-en-auge/

Sistemas de visión por computadora para el control de calidad. (2015). https://www.milenio.com/opinion/varios-autores/universidad-politecnica-de-tulancingo/sistemas-de-vision-por-computadora-para-el-control-de-calidad