Loading...

Moving Object Detection Using Background Subtraction Algorithms

Master's Thesis 2014 58 Pages

Computer Science - Miscellaneous

Excerpt

INDEX

Acknowledgement

Abstract

List of Figures

List of Tables

1. INTRODUCTION
1.1 OBJECTIVE
1.2 APPLICATIONS
1.3 LITERATURE SURVEY
1.4 ORGANIZATION OF THE REPORT

2 BLOCK DIAGRAM AND CHALLANGES
2.1 GENERAL STEPS FOR OBJECT DETECTION
2.2 CHALLENGES

3 STUDY OF DIFFERENT BACKGROUND SUBTRACTION
3.1 SIMPLE BACKGROUND SUBTRACTION METHOD
3.2 MEAN FILTERING METHOD
3.3 MEDIAN FILTERING METHOD
3.4 W4 SYSTEM METHOD
3.5 FRAME DIFFERENCING METHOD
3.6 RUNNING GAUSSIAN AVERAGE MODEL
3.7 GAUSSIAN MIXTURE MODEL
3.8 EIGENBACKGROUND

4 COMPARISION OF BACKGROUND SUBTRACTION

5 OPTICAL FLOW
5.1 THE SMOOTHNESS CONSTRAINT
5.2 DETERMINING OPTICAL FLOW USING HORN - SCHUNCK
5.3 ESTIMATION OF CLASSICAL PARTIAL DERIVATIVES
5.4 EXPERIMENT RESULTS

6 COMBINE GMM & OPTICAL FLOW

7 SHADOW DETECTION
7.1 HSV/HSI MODE
7.2 SHADOW DETECTION

8 CONCLUSION AND FUTURE WORK

REFERENCES

LIST OF FIGURES

1.1 Various Techniques For Object D&T

2.1 Block Diagram Of Object Detection

3.1 Block Diagram Of Simple Background Subtraction Method

3.2 Original Video Sequence

3.3 Simple Background Subtraction with Threshold 10

3.4 Simple Background Subtraction with Threshold 20

3.5 Block Diagram Of Mean Filtering Algorithm

3.6 Mean Filtering With Mean Of 5 Frames

3.7 Mean Filtering With Mean Of 15 Frames

3.8 Block Diagram Of Median Filtering Algorithm

3.9 Median Filtering With Mean Of 5 Frames

3.10 Median Filtering With Mean Of 5 Frames

3.11 Block Diagram Of W4 System

3.12 Output Of W4 System

3.13 Block Diagram Of Frame Differencing

3.14 Output Of Frame Differencing

3.15 Basic Flow Chart Of GMM Algorithm

3.16 Representation of history of a pixel in k=5 distributions

3.17 B=3 distribution with largest weight

3.18 Output Of GMM With Alpha 0.05, SD 9 and Threshold 0.3

3.19 PCA for Data Representation

3.20 PCA for Dimension Reduction

3.21 The PCA Transformation

3.22 Output Of Eigenbackground

5.1 Optical Flow Constrain

5.2 Partial Derivatives Of Image Brightness At Point (X, Y)

5.3 BFB Kernel

5.4 BFB Kernel Apply For X Direction Displacement

5.5 BFB Kernel Apply For Y Direction Displacement

5.6 BFB Kernel Apply For Time Displacement

5.7 An Illustration Of The Iterative Minimization Process For U Where Steps 2 And 3 Are Iterative

5.8 Foreground Detection Using HS Algorithm For Optical Flow

6.1 Block Diagram Of Combined Approach

6.2 Output Of Combined Method Compare With GMM Method

7.1 HSV Mode

7.2 Double cone model of HSV mode

7.3 Output Of Shadow Detection

LIST OF TABLES

illustration not visible in this excerpt

CHAPTER 1 INTRODUCTION

1.1 Objective

Detection of motion is very essential and common step in many surveillance system applications. In these applications, the goal is to obtain very high sensitivity in detection of moving objects with less possible false alarm rates. Background segmentation is one of the famous technique that is commonly used. It works on the intensity difference of the current frame with the background frame. Moving object detection and tracking (D&T) are important steps in object recognition, context analysis and indexing processes for visual surveillance systems etc.. It is a big challenge for researchers to prepare algorithms on which detection and tracking is more suitable for background situation, environment conditions and to determine how accurately object D&T (real-time or non-real-time) is made. There are variety of object D&T algorithms and publications available. On that basis we can compare their performance and evaluate via performance metrics. This report provides a systematic review of these algorithms, their brief description and performance analysis and effectiveness in terms of execution time and memory requirements. When the camera is fixed and the number of targets is small, objects can easily be tracked using simple methods. Computer vision-based methods often provides non-invasive solution. Their applications can be divided into three different groups: Surveillance, control and analysis. The object detection and tracking (D&T) process is a necessary requirement for surveillance applications. The control applications, which uses some parameters to control motion calculations and estimation. These parameters are used to control the relevant vision system. The analysis applications are generally automatic, and used to optimize and diagnose system’s performance. For well predefined (namely, annotated) datasets, the object recognition algorithms give good accuracy.

The aim of motion tracking is to detect and track moving objects using a sequence of images. Motion tracking is not only useful for monitoring activity in public places, but it is becoming a key procedure for further analysis of video imagery. For example, information about the location and identity of objects at different points in time is the basis of detecting unusual object movements or coordinated activities e.g. strategic plays in a football game. Object detection in videos involves verifying the presence of an object in image sequences and possibly locating it precisely for recognition. Object tracking is to monitor an object spatial and temporal changes during a video sequence, including its presence, position, size, shape, etc.

In recent years, with the latest technological advancements, off-the-shelf cameras became vastly available, producing a huge amount of content that can be used in various application areas. Among them, visual surveillance receives a great deal of interest nowadays. Until recently, video surveillance was mainly a concern only for military or large-scale companies. However, increasing crime rate, especially in metropolitan cities, necessitates taking better precautions in security-sensitive areas, like country borders, airports or government offices. Even individuals are seeking for personalized security systems to monitor their houses or other valuable assets. So, our objective is to develop some sort of algorithm that will satisfy human requirements in future.

1.2 Applications

-Traffic monitoring like counting vehicles, detecting and tracking vehicles.
-Human action recognition like run, walk, jump etc. Human-computer interaction
-Object tracking
-Many other computer vision application like digital forensics.
-Optical motion capture
-Content based video coding

1.3 Literature Survey

This survey paper reviews briefly research works on object detection and tracking in videos. The definition and tasks of object detection and tracking are first described, and the potential applications are mentioned. Followed is the summation of major research highlights and widely used approaches. In this literature we study about various techniques of moving object detection and tracking Reference is included at the end of the report.

illustration not visible in this excerpt

Fig.1-1 Various Techniques For Object D&T

Here in this thesis, we are concentrating on moving object detection techniques using background subtraction algorithms [1-5] like Simple Background Subtraction, Mean and Median Filtering. W4 System, Single Gaussian Model, Gaussian Mixture Model and Eigenbackground, their performance and comparison analysis. We have also studied optical flow technique like Horn-Schuck & Lucas-Kanade methods and then by combining it with background subtraction algorithm we can improve our performance specially against false detection due to illumination changes.

1-4 Organization of the Report

This report consists of eight chapters in total. The framework for the report is described as follows:

Chapter 1 provides a brief introduction about the basic vision based video surveillance system and the need for an implementation of this report.

Chapter 2 explains the basic concept and principles of the Background Subtraction Algorithm and how the basic operation done to find moving object. It also list some challenges.

Chapter 3 discusses the widely used Background Subtraction Algorithms and it's description, implementation and output analysis .

Chapter 4 gives comparative analysis of methods describe briefly in chapter 3 based on their speed, memory requirement and performance.

Chapter 5 explains the basics of optical flow theory with brief of standard HS algorithm, its procedure to obtain optical flow and its output.

Chapter 6 presents the implementation of combine approach of GMM and Optical Flow and its result comparison with standard GMM output.

Chapter 7 presents the basic HSV threshold approach for shadow detection.

Chapter 8 Finally, I give the reader an overall summary on the collected results and future work.

CHAPTER 2 BLOCK DIAGRAM AND CHALLANGES

This chapter contains the basic flow, which is require for the implementation of object detection and tracking. Though these steps can be implemented using many techniques. Also, this chapter explains what are the most common challenges and problems, which must be taken care for good and acceptable results.

2.1 General Steps For Object Detection

To perform high level tasks like Video Surveillance, traffic monitoring and Automated event detection, it is necessary to model the background. The performance of these systems is depended on the accuracy and speed of object detection algorithm.

illustration not visible in this excerpt

Figure 2-1 Block Diagram Of Object Detection

The basic block diagram of object detection procedure is explain in the figure 2-1. The real time image is captured through camera and then using background subtraction algorithm by having a reference background, we can achieve noisy foreground. This noise can be reduce by applying filtering operation, generally Morphological Operation is carried out in this phase. After obtain connecting region, we can able to find foreground mask and ultimately by applying this mask we can detect and track our object. For real time processing Background subtraction algorithm is well suited.

2.2 Challenges

There are many challenges to deal with when we have to implement Background Subtraction Algorithm as listed below.

Illumination changes like clouds moving in the sky Motion changes like swaying of trees Secondary illumination effects like static shadows Moving Shadows and camouflage.

CHAPTER 3 STUDY OF DIFFERENT BACKGROUND SUBTRACTION ALGORITHMS

In this chapter we have discussed many commonly used Background Subtraction methods which are highly used for segmentation of given image in order to find foreground mask.

3.1 Simple Background Subtraction Method

In basic method for Background subtraction, the static background image without object is taken first as a reference image. After that the current image of the video is subtracted pixel by pixel from the background image and resultant image is converted into binary image using threshold value. This binary image is worked as a foreground mask. For conversion in binary image threshold is required. From1 we can write

illustration not visible in this excerpt

Where, ܫ௧ሺݔǡݕሻ is pixel intensity of frame at time t, B(x,y) is mean intensity on background pixel and T is threshold. When difference reaches beyond threshold the pixel categorize as a foreground pixel.

So the effectiveness of the object detection is depends on the threshold value.

Although this method is very fast, it is very sensitive to illumination changes and noise.

illustration not visible in this excerpt

Figure 3- 1 Block Diagram Of Simple Background Subtraction Method 7

Figure 1- 2 Original Video Sequence

Figure 3- 3 Simple Background Subtraction with Threshold 10

Figure 3- 4 Simple Background Subtraction with Threshold 20

For result analysis we have taken two video sequences6 from which some of the frames are shown in Figure 3-2. After taking consideration of reference frame we are subtracting each frame from reference frame. Then, we can use threshold T as per equation (1) to obtain binary image as shown in figure 3- 3 for Threshold value 10 and in figure 3-4 for threshold value 20. We can also apply dilution and erosion operation to improve our foreground mask. From comparing results from two threshold we can say that the output mask is very much sensitive to threshold value. Though, this algorithm is fast but also very much sensitive to noise and lightning changes.

3.2 Mean Filtering Method

The next method is mean filtering. Figure 3-5 gives as the basic functionality of Mean Filtering. In this method the reference frame is calculated using the mean of the last n frames as per equation 2 given below.

illustration not visible in this excerpt

(2)

where, B(x,y,t) is reference background calculated at time t and I(x,y) is the pixels intensity. So, the foreground can be found using equation 3, where T is a threshold.

illustration not visible in this excerpt

(3)

illustration not visible in this excerpt

Figure 3- 5 Block Diagram Of Mean Filtering Algorithm

Chapter 3 : Study Of Different Background Subtraction Algorithms

Figure 3- 2 Mean Filtering With Mean Of 5 Frames

Figure 3-3 Mean Filtering With Mean Of 15 Frames

As shown in figure 3-6 & 3-7, we have calculated mean image from last 5 & 15 frames and in order to obtain mask we have subtracted mean image from ith image. we have consider T = 8 for both case. It is easy to implement and adaptive for background calculation. But accuracy is depends on object speed, also memory requirement is very high and have global threshold i.e. same threshold for every pixel. Also from results we can comment that if training signal is too long than memory requirement is very high and if to short than updating rate gets very high. So, slow moving object is not detected property.

3.3 Median Filtering Method

illustration not visible in this excerpt

Figure 3- 8 Block Diagram Of Median Filtering Algorithm

The another method is median filter. Figure 3-8 gives as the basic functionality of Medeiann Filtering, in which background model B(x,y,t) is given by equation 4.

illustration not visible in this excerpt

(4)

The advantages of the these two above methods are they are fast, easy to implement and adaptive background calculation. The disadvantages are accuracy is depends on object speed, also memory requirement is very high and global threshold i.e. same threshold for every pixel.

As shown in figure 3-9 & 3-10, we have calculated median image form last 5 & 15

illustration not visible in this excerpt

Figure 3-9 Median Filtering With Mean Of 5 Frames

illustration not visible in this excerpt

Figure 3-10 Median Filtering with Mean Of 15 Frames

frames with T= 8 and in order to obtain mask we have subtracted mean image from ith image. It is easy to implement and adaptive for background calculation. But accuracy is depends on object speed, also memory requirement is very high and have global threshold i.e. same threshold for every pixel. Also from results we can comment that if training signal is too long than memory requirement is very high and if to short than updating rate gets very high. So, slow moving object is not detected property.

3.4 W4 System Method

The next method is based on method used in W4 system. As mentioned in [1-2] and7, we can say, during training sequence maximum intensity, minimum intensity and maximum intensity difference (Dmax) in successive frame of the pixel is calculated and then foreground pixel is calculated based on following equation. Also the basic operation is mentioned in figure 3-11.

illustration not visible in this excerpt

(5)

[...]

Details

Pages
58
Year
2014
ISBN (eBook)
9783656672661
ISBN (Book)
9783656672678
File size
4.4 MB
Language
English
Catalog Number
v275108
Grade
9.2
Tags
moving object detection using background subtraction algorithms

Author

Share

Previous

Title: Moving Object Detection Using Background Subtraction Algorithms