Introduction Of Video Surveillance

Published Date: 02 Nov 2017

CHAPTER 1

INTRODUCTION

1.1 INTRODUCTION OF VIDEO SURVEILLANCE

Object detection is the fundamental aspects in video analysis. Detection of Moving object is a difficult task in surveillance systems. Video surveillance is a process of analyzing video streams. Although many works aimed at detecting objects in video streams have been reported. Due to fast illumination changes in surveillance system, many are not suitable for dynamic background. It is an active area in computer vision. There are three types of Video surveillance activities. Video surveillance activities can be manual, semi-automatic and fully-automatic.

Video content is analyzed by the human in manual video surveillance system. Manual video surveillance system is widely used. In Semi-automatic video surveillance system, video is analyzed by significant human intervention. Only in the presence of significant motion the video is recorded and sent for analysis by a human expert. By a fully-automatic surveillance system, the input is the video sequence taken at the scene where surveillance is performed. In this system human effort is not needed and the system does low and high level tasks such as object detection and tracking, unusual event detection and human action recognition. The video surveillance system that supports automated objects classification and object tracking.

1.1.1 MOVING OBJECT DETECTION

Moving object detection is the challenging task in video surveillance system. Moving object detection consists of static background and dynamic background. In static background, optimal flow method and inter frame difference method are commonly used for background subtraction to detect the objects. In the dynamic background, motion compression and segmentation method are commonly used to detect the objects.

1.1.5 Static Background

(i) Background subtraction

The background elimination method is used to detect a motion object in the video surveillance area by using the variance of the current and the background image.

(a) The initialization and updating background

In the static background, video streams of the background image is responsive to the effects of the weather, illumination, shadow and other interference, so the background model wants to be updated. The parameter model and statistical estimation method are the two methods to initialize and update the background. To establish the background in parameter model, the Gaussian distribution is used and it adjusts the parameters adaptively. Thus a new background image will be obtained. A series of images is used in statistical estimation method and the pixel gray is classified according to certain assumptions, and the appropriate pixels are chosen to form the background. The adaptive background method and margined sign correlation being the two approach used in the Graphics Processing Unit. This approach detects moving objects and take off shadow regions more effectively.

(b) Shadow detection and removal

Due to the influence of light, the shadow of the moving target is often detected as moving targets, leading to the misjudgment of the tracking system, so the shadow should be removed. Shadow detection and removal methods are used, first method is based on shadow feature and the other method is based on the geometric model. The main aspects of the shadow detection are texture, color, and geometry features. The features of the shadow are used to detect a moving region and again it separate the object from the shadow. HSV color space is performed well to eliminate the shadow, but the changes of hue and saturation values should be within a certain range. This method is widely used. In geometric model, the geometric features of the camera positions and scene surface are used to detect shadows. This method needs prior knowledge.

(ii) Optical flow method

The moving target is detected through the study of optical flow field of the video streams. This field consists of important information of the moving target. Optical flow model is the motion vector of the pixel in the image grayscale mode. The key step of the optical flow method is to obtain the optical flow information of the moving target. In the optical flow field calculation, the movement of the object is typically continuous in space.

The optical flow method is sensitive to noise and easy to produce incorrect results. The inter-frame difference method, first wants to obtain a moving target area and then want to use the optical flow method to obtain accurate target detection. The low accuracy of the frame difference method and the demerits of the this field can be avoided by this approach.

(iii) Frame difference method

The frame difference method is calculated by the variation between two or three consecutive images, and then the information of moving objects can be obtained. This method is simple and self adaptive to the dynamic environment. The limitations in frame difference methods are easy to produce small holes and it cannot detect the complete target. To improve the performance of the frame difference method, the fast symmetric difference method is used and it can be used to detect the shape of the moving object more accurately. This method work well in noisy environment with robustness.

1.1.6 Dynamic Background

In dynamic background, the global motion region consists of background and the local motion region consists of a foreground. The global motion is the overall action of all of the pixels in the sequence of images. There are three methods to detect moving object in dynamic background, they are motion compensation method, motion segmentation method, regional integration method.

(i) Motion compensation method

The motion compensation is a method to describe the difference between adjacent frames. First step in this method is to calculate the movement. The second step is static background detection algorithm is used to identify the moving object. The motion compensation method is divided into two major categories. The first method is block motion compensation, each frame of the video sequence is divided into a number of pixel blocks and the current block is predicted and also to be compensated according to the reference block.

(ii) Motion segmentation method

This method is used to accomplish the moving target segmentation by calculating the motion vector of the each pixel in the horizontal direction. The segmentation based on the optical flow and the change detection method, parameterization and Bayesian segmentation method are included.

(iii) Regional integration method

The moving target is segmented and it makes use of the visual features and then the motion parameters are calculated. The image segmentation methods based on wavelet transform and feature weighting, which can segment random texture image.

1.1.2 Applications of surveillance system

Visual surveillance is an exciting research area among all the areas of image processing system. It's especially used for detecting humans and vehicles in dynamic scenes.

Surveillance Applications are:

(i) Access control in special areas.

In highly secured areas decided by authority agencies needs to be restricted only to authenticated people. Such as military bases and important governmental units, only authenticated people allowed to enter the premises. In order to implement this restriction a comprehensive biometric solution which determines the parameters like height , facial expression needs to be authenticated. It should be automatically detected and compare with the legal visitors those are enrolled in the system.

(ii) Person-specific identification in some scenes.

Personal identification with a surveillance system nab the mistrust at public locations. It requires the database of suspects and a smart video surveillance system. The police can keep the distance where the culprits move in. This kind of biometric solution can minimize the human error and improve the accuracy of system prediction.

(iii) Crowd flux statistics and congestion analysis.

One of the befitting scenarios of video surveillance systems in modern life is a congestion analysis. In modern transport system a dynamic solution that predicts and live traffic analysis is the cornerstone of an effective system. This biometric system requires to monitor road network and live traffic analysis.

(iv) Anomaly detection and alarming.

In a modern decision system requires to build upon video surveillance system one such application is anomaly detection in public place. It requires to analyze the movement of people and detect the anomaly in behavior. This kind of solution can be set in parking areas and supermarkets in time it raises alarm if an anomaly is detected.

(v) Interactive surveillance using multiple cameras.

For security in social places requires an interactive muiticameras solution. It can be used to track suspects from feeds that come from multiple camera. For traffic management police requires track and catch the vehicle that involved in affording traffic.If anything detected any rules is violated it automatically detect the vehicle number and take affordable solution.

1.2 PROJECT DESCRIPTION

Video analysis is the common platform that assimilates the information from the camera to provide security in various scenarios. It is used in military or security environments, industrial applications, Health care applications and shopping mall areas etc. video analysis task is important in everyday life. Information taken from the video cameras is still error prone to human performance.

In video analysis, segmentation is the basis part for detecting moving objects in the video sequence. The proposed model consists of SOM to work in dynamic background for segmentation. Detecting the moving object using the SOM in video streams are not suitable for dynamic background and it requires complex computation to adjust the threshold values based on Hue Saturation Value (HSV). The SOM consists of an unsupervised learning algorithm. The proposed model also includes a Fuzzy Extreme Learning Machine, it automatically determines the parameter values required to process different video sequences with less human intervention. This method preserves the robustness in its performance and this method works well in dynamic background.

CHAPTER 2

LITERATURE SURVEY

The main aim to perform the background Elimination is to determine the moving object from the video streams in surveillance system. Identifying the moving objects in a video sequence is the potential task in many computer environments. There are many difficulties in First, it must be robust against changes in illumination. Second, it should avoid detecting non-stationary background objects such as moving leaves, rain, snow, and shadows cast by moving objects.

Video surveillance systems have grown steadily in recent years, but most are not tolerant to the dynamic background. Differencing of adjacent frames in a video sequence has been used for object detection in stationary cameras. However, it was realized that straightforward background subtraction was unsuited to surveillance of real-world situations. Any motion detection system based on background subtraction, needs to handle a number of critical situations such as: gradual variations of the lighting conditions in the scene, small movements of non-static objects such as tree branches and bushes blowing in the wind, noisy image due to a poor quality image source, permanent variations of the objects in the scene, such as cars that park, sudden changes in the light conditions, multiple objects moving in the scene both for long and short periods, shadow regions that are projected by foreground objects and are detected as moving objects. The difference between the current image and the background is used to detect the motion. It does well in motion detection when the background model is good. So for these algorithms, the background model is the key. Most background modeling approaches tend to fall into the category of pixel-wise models.

Background modeling approaches consist of two steps: the proper updating of a reference background model, and the suitable subtraction between the current image and the background model. A simple adaptive filter has been used in to update recursively the statistics of the visible pixels. In the Kalman filter is used to model adaptively the background pixel according to known effects of the weather and the time of day on the intensity values. In color and edge information has been used both for background modeling and for subtraction, using confidence maps to fuse intermediate results. In recent years, a lot of methods are proposed to segment moving objects in dynamic scenes Gaussian mixture model (GMM) is one of the most popular models for moving object segmentation, which model the color of every pixel in the image with a mixture of Gaussians model. With regard to GMM, some similar methods are proposed. To handle the sharp changes of illumination in scenes, a hierarchical Gaussian mixture model is used in background modeling.

Independent component analysis algorithm is used for background subtraction to detect the moving and motionless persons in indoor surveillance system quickly. It is used to detect the object in the Indoor scenes and applications such as Homecare and healthcare monitoring system. In an independent component analysis, the background subtraction consists of two stages. Two stages are training and detection. In the training stage, ICA models directly measure the statistical independence based on the estimations of joint and marginal probability density function. On the detection stage, the trained vector is used to distinguish the foreground image with respect to the reference background image.

The HSV technique is used for detecting the shadow and suppression in a system. HSV color space is used to improve the performance in detecting shadows, because moving shadow can affect the current localization and detection of moving objects.  The neural Fuzzy technique is used to detect the object in the dynamic background without any human intervention. on the neural stage, SOM is used to detect the objects and eliminates the shadow by using the HSV color space. On fuzzy stage,sugeno and the mamdani fuzzy system are used. By using that fuzzy methods, video segmentation is done. The drawbacks in the sugeno fuzzy system are more complex and it takes more computation time and it also requires more memory to compute the parameters.  

An adaptive Local patch Gaussian Mixture Background Model is used to detect the moving objects from video with dynamic background. Then the SVM classification is used to distinguish between foreground objects and shadow areas. The Fuzzy Extreme Learning Machine and it automatically determines the parameter values for the given Input frames. It randomly chooses and fixes the hidden node parameters and then analytically determine the output weights. This method is well suited to work in dynamic background.

An adaptive background model is based on statistical probability and shadow suppression scheme and is used to detect the both indoor and outdoor scenes. This method uses the HSV color space for shadow detection. This algorithm is more efficient and robust.

An efficient method is used for detecting ghost and left objects in video surveillance system. This method consists of two main stages, one stage is to detect the stationary objects and second stage is to discriminate the conditions between ghost and left objects. It preserves the robustness in performance.

Bayesian foreground and shadow detection is used to detect the objects in dynamic background. Finally Markov random field model is used to increase the accuracy of the separation. This algorithm works well in indoor and outdoor scenes.

Parametric and nonparametric model

Parametric models are tightly coupled with underlying assumptions, not always perfectly corresponding to the real data, and the choice of parameters can be cumbersome, thus reducing automation. The nonparametric models are more flexible but heavily data dependent.

Unimodal and multimodal

Basic background models assume that the intensity values of a pixel can be modeled by a single unimodal distribution. Such models usually have low complexity, but cannot handle moving backgrounds, while this is possible with multimodal models at the price of higher complexity.

Recursive and nonrecursive

Nonrecursive techniques store a buffer of a certain number of previous sequence frames and estimate the background model based on the temporal variation of each pixel within the buffer. The recursive techniques recursively update a single background model based on each input frame. In the first case, the background well adapts to eventual variations, but memory requirements can be significant.

Pixel-based and region-based methods

Pixel-based methods assume that the time series of observations is independent at each pixel, while region- based methods take advantage of inter pixel relations, segmenting the images into regions or refining the low-level classification obtained at the pixel level. This increases the overall complexity. Our approach is based on the background model automatically generated by a self-organizing method and can be broadly classified as nonparametric, multimodal, recursive, and pixel based.

CHAPTER 3

PROBLEM STATEMENT AND ITS SOLUTION

3.1 EXISTING SYSTEM

A fundamental aspect in video analysis is related to the segmentation of moving objects in video sequences. Once this problem is solved, the huge predicted benefits of video analysis can be reaped. There are many segmentation techniques are present. Some of the segmentation techniques are used and but many are not tolerant to the dynamic background.  The neural Fuzzy technique is also used to detect the object in the dynamic background without any human intervention. on the neural stage, SOM is used to detect the objects and eliminates the shadow based on HSV color space. On fuzzy stage,sugeno and the mamdani fuzzy system are used. By using these fuzzy methods, video segmentation is done. The drawbacks in the sugeno fuzzy system are more complex and it takes more computation time and it also requires more memory to compute the parameters. The mamdani fuzzy system is easy to use.  

Disadvantage of Existing System

Lacks robustness.

Complex computation.

Addition to manual parameter adjustments.

3.2 PROPOSED SYSTEM

The proposed Model aimed to detect the moving objects from a video streams in Dynamic Backgrounds with stationary cameras. The Self Organizing Map (SOM) decreases the computational load to improve the video segmentation. During the training phase, the output provides the highest activation unit to the given input pattern and declared as the winner. The output node whose incoming weight is the shortest Euclidean distance from the input vector is considered as the winner-take-all. The process of the SOM is to select the output layer topology and train the weights from input to output layer. The proposed model also includes the Fuzzy Extreme Learning Machine, it reduces the time to train neural network and classification is done with high accuracy.

Advantage of Proposed System

Quick Learning Speed.

Good generalization performance.

Decision making is fast.

The method preserves the robustness in its performance.

It need not require previous training.

CHAPTER 5

SYSTEM DESIGN

5.1 OVERVIEW OF MODULES

Background elimination

Moving object detection in Dynamic Background

Extreme Learning phase

1. Background elimination

The main goal of background elimination is to detect the objects from the portion of a video sequence of a fixed camera. The background contains static scene. Foreground contains moving objects.

2. Moving object detection in Dynamic Background

Detecting the moving object using the SOM in video streams are not suitable for dynamic background and it requires complex computation to adjust the threshold values based on Hue Saturation Value (HSV). By using HSV color space, Shadow is removed effectively.

3. Extreme Learning phase

Fuzzy-Extreme Learning Machine (FELM) is used for detecting the object in dynamic backgrounds. The Fuzzy Extreme Learning Machine contains Single Hidden layer Feed- forward neural networks (SLFNs). Extreme Learning Machine randomly choose the weight for the input and analysis the output weights. This model is suitable for Dynamic Environment. The proposed model involves Fuzzy-Extreme Learning Machine and Self Organizing Map (SOM) which are used to detect the moving objects as well as shadow elimination in dynamic background.

5.2 FLOW DIAGRAM

Figure 1: Flow diagram of Proposed Method

CHAPTER 6

IMPLEMENTATION

6.1 DESCRIPTION OF MODULES

6.1.1 Background elimination

The main goal of the background subtraction is to detect the moving objects in the video streams taken from the stationary camera. The background contains static scene. Foreground contains moving objects.

6.1.2 Moving object detection in dynamic background

The proposed model deals with dynamic background situations. The video samples used in our system are color videos. The video sequence 1 is taken from the webcam and it can have the resolution of 120×160pixels at 15 frames per second. The video sequence 2 contains indoor scenes and has resolution of 320×240 at 24 frames per second. In this proposed model, video is given as the input and that video are converted into frames. The main aim of the background elimination is to detect the moving objects from the video sequence taken from the stationary camera. In SOM, for each video streams want to adjust the manual threshold value e1 and e2. The e1 and e2 is the segmentation threshold. Finally, the dynamic moving object is detected. It preserves the robustness in its performance.

In this model, input is given as the video file in AVI format. First step is to separate the frames from the video. Then convert it from RGB plane to HSV plane and it forms the neuronal map for the first frame (HSV) and that frame is considered as the background. The HSV components mapped with each pixel (x, y) of the first video frame and update weights. Each neuron is connected only to its similar pixel and each pixel is represented in the HSV color space. Finally, the remaining frames will be converted to HSV and it will be compared to the background and detect the moving objects from the remaining frames by using the values e1 and e2. The e1 and e2 segmentation threshold value. The shadow detection is performed well in Hue-Saturation-Value (HSV) color space..

The Euclidean distance between the two pixels Pi=(Hi,Si,Vi) and Pj=(Hj,Sj,Vj) based on HSV color hexone is calculated as D(Pi,Pj)=||(ViSicos(Hi),ViSisin(Hi),Vi)-(VjSjcos(Hj),VjSjsin(Hj),Vj)||22

Figure 2: Single frame from each video sequence

Procedure

Step 1: Convert the video sequence into a number of frames.

Step 2: Set first frame as a background.

Step 3: Calculate HSV for the frame and form Neuronal map ‘NM’ using weight vectors.

Step 4: Calculate HSV for the next frame.

Step 5: Calculate Euclidean Distance between HSV and Neuronal map.

Step 5 (a): Set the segmentation threshold value e1 and e2.

Step 6: If Minimum distances D<e1, Set the corresponding pixel as background.

Step 6 (a): Else the distance D≥e2, Set the corresponding pixel as foreground.

Step 7: Update the weight vectors.

Step 8: Else set as object.

Step 9: Repeat Step 4 to Step 8 up to the last frame.

6.1.3 Fuzzy Extreme Learning Machine

The Fuzzy Extreme Learning Machine uses Single Hidden layer Feed- forward neural networks (SLFNs). It automatically determines the threshold value for the given video sequence to detect the moving object. This model is appropriate for Dynamic Environment. SLFNs randomly choose the weight for the input and analysis the output weights. The SLFN have n pairs of approximate input and output values and also have P hidden nodes namely, z= (xi,yi) where xi  Rn, yi  Rm, for i = 1, 2, …, n, then standard SLFNs with P hidden nodes and output function k(x) are modeled as βiS(ai,xj,bi)=tj, where j=1,2,….,n.(ai,bi) are the hidden node parameters and βi are the weight vector .The βi connects the ith hidden node and the output node. The mathematical model is given as βiS(ai,xj,bi)=tj, where j=1,2,….,n and it is equivalent to Hβ=T.

h(x1) S(a1,x1,b1) …………S(ap,x1,bp) .

H= . = . .

. . .

h(xn) S(a1,xn,b1) …………S(ap,xn,bp) n×P

β1T t1T

. . . β= . and T= .

βPT P×m tpT n×m

The output matrix H is calculated for hidden layer. The output of the ith node is from ith column of H with respect to input x1,x2,….,xn.The output weight vector is β = H# T. H# is the Moore-penrose generalized inverse of Hidden layer H.

ELM Algorithm

For the given set of training input / output values {z= (xi,yi) where xi  Rn, yi  Rm, for i = 1, 2, …, n}, the activation function k(x) and the hidden nodes P.

Step 1: Generate random hidden node parameter (ai,bi) for i = 1,2, …, n by using continuous distribution.

Step 2: The output matrix H is calculated for hidden layer.

Step 3: The output weight β is calculated based on the connection β = H# T.

CHAPTER 8

RESULT AND ANALYSIS

Fuzzy-Extreme Learning Machine and Self Organizing Map (SOM) which are used to detect the moving objects as well as shadow elimination in dynamic background. The moving object detection is determined by using the value of e1 and e2 generated by the human and with FELM. The e1 and e2 are the segmentation threshold value. The original video is converted into frames and the length of the stream is displayed. The original frame is given as the input. By using SOM and FELM, the manual parameter is adjusted by a human viewer for each video sequence and they detect the moving object and eliminate the shadow more effectively.

ANALYSIS

TABLE I

e1 and e2 values determined by Human Viewer

Video

e1

e2

1

0.06

0.05

2

0.10

0.08

The video sequence 1 containing outdoor scenes. The corresponding foreground is calculated by SOM and FELM. In SOM, the human assigns the parameter values of e1, e2, L1, L2. e1 and e2 are the segmentation threshold values. The c1 and c2 are the learning rate to measure and update the background quickly. For video sequence 1, the foreground value calculated by SOM are e1=0.06, e2=0.05 as shown in Table I. Finally the moving object is detected. The video sequence 1 containing outdoor scenes and the video sequence 2 containing Indoor Scenes. The FELM automatically determine the parameter value and determine the moving object more effectively. Likewise the other video sequences are computed and it determines the moving object in surveillance system. The FELM randomly generating training and testing dataset. The average training accuracy is 0.9733. The average testing accuracy is 0.9496. By using this FELM method, it reduces the time to train the neural network.

CHAPTER 9

CONCLUSION AND FUTURE ENHANCEMENTS

The proposed method is the automatic model, it doesn’t require previous training. It automatically tuned the threshold values for each video without any previous training. By using SOM and FELM, the moving objects are detected in dynamic backgrounds and shadows are eliminated more efficiently. This proposed method reduces the time to train and test the neural network with high classification accuracy. This method works well in dynamic environments. Detecting the moving objects in video sequences is the interesting problem. Future work will address some techniques like frame differencing, SVM classification to get better results in moving object detection in video surveillance system.