Face Detection and Tracking in a Video Sequence

Guided By:Ranganatha S B.E,M.Tech,MISTEAssistant Professor

Presented By:4GH12CS019 Karthik G NDepartment of Computer Science & Engineering, Government Engineering College, HassanMay-2016Face Detection and Tracking in Video SequenceFinal phase project evaluation on....

IndexIntroduction ProblemSolutionArchitecture and DesignProject Result AnalysisChallenges and ConstraintsConclusion and Future WorksReferences

IntroductionVideo processing has become a major requirement in current world.This technique is majorly used to detect, recognize and track various objects.Face detection and tracking is the phase where we detect a persons face from a video sequence and track him/her throughout the video.It plays vital role in video corrections, surveillance, military tracking so on.

Problem There are many existing algorithms for face detection and tracking in video sequences. But none of them have an accuracy of tracking the facial region completely.There is no algorithm till date that tracks all kinds of facial features in videos under all possible constraints effectively.

SolutionDeveloping a modified algorithm from existing algorithms to increase the accuracy. The increase in tracking accuracy is achieved by fusing two different algorithms that work based on similar concepts and similar point of interest. The new fused Face detection and tracking algorithm provide more accuracy due to the fact that it combines two algorithms, it is a simple logic that if one algorithm fails to track the facial region, other algorithm keeps track of it and gradually the accuracy will be improved.

Architecture and Design

Modules Design

ProjectOur project begins with the detection of face in the 1st frame in the video sequence using Viola-Jones Algorithm.We used the Viola-Jones detector to detect face in the input video sequence using MATLAB Toolbox.Output of the detector is fed as a input to masking, masked in such a way that the rest area apart from the face region in the 1st frame is masked out.We obtain the ROI ( face region in our case ) in the frame.We convert the input RGB frame to Gray frame using rgb2gray() function.

Continued...We apply Gaussian filter on the computed values.After processing all these steps we apply the Sobels edge detector Algorithm on the modified frame.We henceforth obtain all the computer distinguishable edges in the ROI of the 1st frame.By using these points we find the centroid in the ROI.Tracking starts by calling external function named next2().Tracking uses point tracker to track the points in facial region of the frames.The edge points are concatenated with Eigen Feature points.

Continued...The new concatenated points matrix is fed to the point tracker of KLT algorithm.These points are tracked till last frame of the video sequence that has been given as input.After completion of tracking, the number of frames that contain bounding box is calculated. The resulting value is compared with that of the value obtained by tracking the same video sequence in KLT algorithm and results are tabulated.

Result Analysis

Snapshots

GUIDetected Face

Contd..

Detected PointsTracking Face

Challenges and ConstraintsThe face must be present in the first frame of the input video sequence.The video must be recorded only by fixing the camera in one particular location or fixing the person location and varying the camera.Variation in camera position must be negligible, failure in which leads to increase in complexity while detection and tracking of the faces in video sequence.The input video must be one among many of standard formats used worldwide, change in which leads to false results. As the project fuses various algorithms to increase its efficiency, output binds with the few of the limitations of each algorithms even after overcoming most of their drawbacks.The resulting system must have only one face detected in the first frame, in case there are multiple faces detected then the Sobel's algorithm detect edges but computation of centroid fails leading to failure in tracking of face(s) in further frames of the video sequence.

ConclusionWe have developed a fused Face detection and tracking system which works based on the point tracking as that of KLT algorithm. From the test reports we could clearly observe that fused FDT algorithm tracks face in few more frames than KLT algorithm alone would have achieved and also because we use centroid as one of the point while tracking, the chances of variation in bounding box size and shape is very negligible compared to KLT algorithm alone.

Future WorksModify Viola-Jones algorithm to remove the constraint of face being present in first frame itself.Faces can be detected in further frames using a loop.Generating more points using mid-point theorem from edge points.Eliminating the use of Eigen features for tracking, using point tracker only for the edge points and other generated points.Reducing the execution time by simplifying the code statements.

Referenceshttp://in.mathworks.com/products/image/index.htmlhttp://in.mathworks.com/help/images/http://www.tutorialspoint.com/dip/http://in.mathworks.com/academia/students.html?s_tid=acmain sp_gw_bodhttp://in.mathworks.com/help/matlab/creating guis/about-the-simple-programmatic-gui-example.htmlRafel C Gonzalez and Richard EWoods, \Digital Image Processing", 3rd Edition, Pearson Education, 2003.Milan Sonka, Vaclav Hlavac and Roger Boyle, \ Image Processing, Analysis and Machine Vision", 2nd Edition, Thomoson Learning, 2001.

THANK YOU!!!!