Automating Face Recognition in Surveillance Videos for Forensic Investigation

Abstract

Face recognition is now evolving into forensic science like other investigation methods. As we know how the evidence are collected using Fingerprint matching or blood stains analysis for which more often DNA matching is carried from the information available at the crime scene. For all these proceedings human interaction is required, and here is where face recognition takes advantage. Mainly in a face recognition process the input given is a face image or a video. Therefore there is no need for the person to be present for the identification and veri cation process. Moreover its useful and accurate when for instance fingerprint data is unavailable. This master thesis deals with real time algorithms and techniques for face detection and face tracking in videos. Image based face recognition using eigen faces is discussed. The Eigenface approach uses Principle Component Analysis for dimensionality reduction of the dataset. Eigenface approach is computationally simple and gave good results in con- straint environment. The sole purpose for the image based face recognition was for the understanding of face detection and tracking in videos. For Video based face recognition, two methods are discussed CAMShift Algorithm which uses the skin tone for the detection of face in the videos and Kanade-Lucas-Tomasi tracker which uses the spatial intensity information for tracking face in the videos. Both the algorithms are implemented and results are shown presenting KLT to be better than CAMShift. A new approach is presented where object segmentation is performed, to identify the re- gion of interests i.e face in the videos. Also Cascade object detector is incorporated into the face detection algorithm which allows the algorithm to update the expected position of the detected faces in the next frame. This continuity between the videos frames was not exploited by the CAMShift and KLT algorithms. Thus, in contrast to CAMShift algorithm and also to the Kanade-Lucas-Tomasi tracker, the proposed face tracker pre- serves information about the near positive and gives better results.