PERSON RE-IDENTIFIC USING MATLTION USING MATLAB
D.DHILIP KUMAR1, K.SAKTHI PRIYA1, A.DEEPAK1
1SAVEETHA SCHOOL OF ENGINEERING, SAVEETHA UNIVERSITY.
Person Re-identification is capable of identifying or verifying a person from a digital image or a video frame from a video source. One of the ways to do this is by comparing selected facial features from the image and a facial database. It is the task of recognizing camera views against the background of similar persons. It is a challenging method and it is perfomed by using Matlab. Viola-Jones algorithm is used in this person Re-identification method. The actual advantage of using Viola-Jones algorithm over the other algorithms is accuracy, speed, clarity and uniqueness. In this paper Viola- Jones, correlation matching and gradient boost algorithm are explained to re-identify the person with the help of Support Vector Machine classifier.
Keywords: - Person re-identification, viola-Jones, gradient boost, correlation matching, SVM
Person re-identification is the computing of identifying a person in a digital image or video. It has many applications in our daily life; most commonly used applications were video surveillance, biometric passwords and social networking sites . The visual surveillance is very important for the public security at the recent years. Surveillance is monitoring the behaviors, activities and other changing information. These can be monitored using the electronic equipment such as CCTV cameras. It is mostly used in the commercial and public security, military, traffic surveillance etc. There are various algorithms like Viola Jones Face Detection Algorithm, Local binary pattern (LBP), AdaBoost Algorithm for Face Detection, SMQT Features and SNOW Classifier Method, Neural Network-Based Face Detection, KLT algorithm, Principal component analysis . Every algorithm has their own advantages and disadvantages, these are used according to the user application. But Viola-Jones algorithm is one of the successful algorithm because of its uncompetitive detection speed and high detection accuracy. So it is one of the successful techniques that have been used in the person identification technique .
2. METHODS FOR PERSON RE-IDENTIFICATION
Figure 1: Block diagram for Person Re-Identification
Person re-identification is performed by dividing the whole process into three stages such as Pre-processing stage, Face Separation stage and Output stage. Each stage performs a specific function.
2.1. Pre Processing Stage
Figure 1.1: Pre processing stage
2.1.1. Input Video
The first and foremost stage in the Person Re-Identification process is the Pre-processing stage. The first step in this stage is to get an input video which contains different persons in that video. This video contains the different person present in the different angle, pose, and lightening condition .
Input video is the collection of many images which composed of moving picture. It consists of many frames presnt in it. In this stage frames present in the input video are separated .
2.1.3. Creating Data Base
The separated frames from the input video are created as the data base for the person re identification techniques. It is used to choose the frame from the collection of separated frames from input video to re-identify the particular person .
Frame which is selected from the data base is resized in this process for the better re-identification method .
2.1.5.Gray Scale Image
Frames which is separated from the input video consists of RBG mixing in it and so we cannot get the clear output to avoid this problem the selected frame is converted in to gray scale which the black and white image.
2.2. Face Re-identification Stage
Figure 1.2: Face Separation Stage
Viola Jones is the algorithm which is used in this method. It is one of the efficient way for the person identification process. It is 15 times quicker than any other algorithm and it is 95% accuracy . It uses the simple haar like feature that are evaluated quickly through the use of the new image representation. The principle of viola Jones algorithm is scanning the detector of same image in different sizes . It is performed with the help of following.
Haar features are similar to convolution kernels which detect the presence of that feature from the image.Each feature will be calculated by subtracting the sum of pixels under white from sum of pixels under black .
In this method value at pixel (x,y) is calculated by sum of pixels above and to left of (x,y). Integral image calculates value of any pixel in the given images .
Adaboost is the one of the machine learning algorithm which helps in finding best feature from different feature. It evaluates and detect the faces from the given pixel window of the input image .
All the feature are cascaded and grouped into many stages. Each stage will have certain number of feature. Each stage concentrates on finding whether the given stage is a face or not.In simple words the integral image used for introducing new image and their features are detected. Adaboost learning algorithm selects the collects the similar feature. Cascade combines all the features and compute to detect required face .
Viola Jones is used effectively in the project to detect the faces from the selected frame which is separated from the input video. It can wisely box all the faces which has clear eyes, nose, lips in the selected frame. The detected faces will be cropped by using face cropping technique, the cropped images will be saved in the newly formed data base folder. Those cropped faces are known as template or feature which will be used for face matching or person re-identification in the project .
2.2.2. Correlation Matching
Digital image correlation is an optical method which is used for tracking and image registration techniques for accurate 2D and 3D measurements of changes in images. When the Frame is selected in the data base, the faces present in that frames are detected using the Viola jones algorithm. To re-idetify the particular person select any particular person images and it re-idetify the particular person in that frame using the correlation matching .
Correlation matching is the technique which is used in the project to match the selected face template in the selected frame. It will compare all the pixels in the frame to match the needed face. It uses the gradient boost algorithm to match the faces. Gradient boost algorithm finds the value of angle, magnitude and gradient of the template and frame. The value of input template will be in the frame in all the pixels and matched value will be shown as identified face .
2.2.3. Gradient Boost
Gradient boost algorithm is the backbone of correlation matching it takes three parameters such as angle, magnitude and gradient to give the value for the template or frame. Image consists of group of pixels to find the angle value of each pixel so gradient boost algorithm is used. Since the images contains many pixels magnitude of the template have to be calculated to match with frame. Gradient is the main parameter which merge the angle and magnitude to find the contrast of the template from the image which has several pixels. Face regions  will have maximum contrast value in the image which helps in creating boundary box and detect the faces from the image. By boosting the gradient contrast value will be increased in turn it increase the contrast value so it helps in getting effective output by minimizing the contrast value of background and maximizing the contrast value of face regions.
2.3. Output Stage
Figure 1.3: Output Stage
2.3.1. SVM classifier
SVM is a support vector machine. It is a learning algorithm used for data classification and to identify the particular data which it is belongs .
SVM helps in training, classifying new data and tuning. This classifier can be trained by the user as required for the application. In the project SVM used to classify the name of the video in which in the input template is belongs and also identified person belongs. Firstly, It works with training the classifier which is the separated frames will be saved in a data base. Each video has separate data base and named as different video names which will be trained in SVM will be used in classifying the video in future. Secondly, SVM can be used in classifying new data, which means new data can be added in future work. Lastly, it can be tuned whenever user needs to modify the classifier. By seeing all these SVM classifier is best classifier which is more flexible than other classifiers .
2.3.2. Output Images
The output image shows the detecting a particular person image from the group of different personality faces. Viola-Jones detects the faces from the video frames; the template will be selected from the data base. Correlation matching and gradient boost algorithm uses it characteristics to match the template in the frame and it is used to match the selected person face. The goal of person re-identification is to select a certain person in one view and to recognize it in the other view. In the work on hand, we assume that we have already detected the persons in both views and we will refer the image of the selected person to as the probe image.
3. RESULTS AND ANALYSIS
STEP 1: INPUT VIDEO
Figure 6.1: Input Video
Video file is given as input in AVI or MP4 format. The first and foremost stage in the Person Re-Identification process is the Pre-processing stage. The first step in this stage is to get a real time input video which contains different persons in that video. This video contains the different person present in the different angle, pose, lightening condition.
STEP 2 : FRAME SEPARATION
Figure 6.2: Frame Separated
Frames is separated from the given input video. Input video is the collection of many images which composed of moving picture. It consists of many frames present in it. In this stage frames present in the input video are separated. The purpose of this step is to prepare the modified video frames by removing noise and unwanted objects in the frame in order to increase the amount of information gained from the frame.
STEP 3: CREATING THE DATA BASE
Figure 6.3: Data Base Created
Data base is created by storing the separated frames from the input video in a folder. The separated frames from the input video are created as the data base for the person re identification techniques. It is used to choose the frame from the collection of separated frames from input video to re-identify the particular person.
STEP 4: RESIZE IMAGE
Figure 6.4: Resized Image
The selected image from the data base is resized as the part of pre processing. Frame which is selected from the data base is resized in this process for the better re-identification method. Enlarging an image generally common for making smaller imagery fit a bigger screen. It is not possible to discover any more information in the image than already exists, and image quality inevitably suffers. However, there are several methods of increasing the number of pixels that an image contains, which evens out the appearance of the original pixels.
STEP 5: GRAY SCALE IMAGE
Figure 6.5: Gray scale converted image
To remove RGB mixing from the input image, we convert it into grayscale image which is black and white image. Frames which is separated from the input video consists of RBG mixing in it and so we cannot get the clear output to avoid this problem the selected frame is converted in to gray scale which the black and white image.
STEP 6: FACE DETECTION
Figure 6.6: Face Detected Image
The faces in the images are detected using Viola Jones Algorithm. Viola Jones is the algorithm which is used in this method. It is one of the efficient ways for the person identification process. It is 15 times quicker than any other algorithm and it is 95% accuracy. The faces present in the frame are detected in different pose, angle and different lighting condition.
STEP 7: PERSON RE-IDENTIFICATION MATCHING
Figure 6.7: Person Re-identification
Person is identified and matched using the selected template Correlation matching and gradient boost is used to match the selected person face. The goal of person re-identification is to select a certain person in one view and to recognize it in the other view. In the work on hand, we assume that we have already detected the persons in both views and we will refer the image of the selected person to as the probe image.
STEP 8: CLASSIFICATION OF VIDEO
Figure 6.8: Classification of video
Using SVM classifier, the type of Video is specified whether it isVideo1, Video2 or Video3. SVM is a support vector machine. It is a learning algorithm used for data classification and to identify the particular data which it is belongs. SVM helps in training, classifying new data and tuning. This classifier can be trained by the user as required for the application. In the project SVM used to classify the name of the video in which in the input template is belongs and also identified person belongs.
In this project, the person re-identification method, which includes two sub-tasks, detection and identification. The research on the person identification technique has been increased in the recent world. A database collected from a video surveillance setting of 3 cameras and real time input video data is taken for analysis. Based on this database, there are many algorithms used in the person identification techniques. Viola-Jones Algorithm is used for evaluating the performance under the person re-identification scenario. Several popular metric learning algorithms for person re-identification have been evaluated. Yet the person identification is one of the difficult task and emerging field using Matlab and it shows that the Viola-Jones algorithm is one of the effective algorithm for face detection because of its simple and fast detection nature. Correlation matching and Gradient boost algorithm used to re-identify the person. Since its most accurate in detecting faces it is used efficiently in several applications like surveillance, biometric password. From the performance, It is conclude that the person re-identification problem is still largely unresolved, thus further attention and effort is needed and presented an effective approach to solve the person re-identification problem in non-overlapping cameras with multiple shot.
...(download the rest of the essay above)