Generative Model Based Video Shot Boundary Detection for Automated Surveillance

Introduction

Video shot boundary detection or video cut detection technique has huge demand in software for various kinds of surveillance system as well as postproduction of videos, as it is the initial step of video image processing specifically applied to content based video retrieval, video summarization and video segmentation. Video shot boundary detection is still a challenging proposition in various types of video analysis applications. A shot can be defined as a similar in nature to a sequence of frames taken in one camera shot, and a scene can be expressed as a collection of one or more consecutive shots that focus on a particular event(s). For an example in a high road no traffic, normal traffic, heavy traffic or any accidental scenarios may be defined as different. In any high road crossing, more than one camera shots showing different cars running in different direction which may be treated as a scene (Dey et al., 2016). There are different types of cuts and transitions between the shots. A hard cut is an abrupt change between two consecutive video image frames and transitions me be defines as with slow changes in brightness. A dissolve occurs when the images of any shots looks less bright and the next consecutive image frames looks brighter, hence the transition is take a superimposition. The problem of video image analysis has been carried out using several techniques ranging from classical probabilistic thresholding to intelligent classification techniques. Based on the types of transitions between two consecutive shots boundaries, it may be categorized into two types: cut (CUT) and gradual transition (GT). Moreover, gradual transition may further be classified into dissolve, wipe and fade out/in (FOI). Video shot boundary detection (SBD) is aimed to identifying two adjacent shots. As far as the benchmark databases are concerned, NIST has a huge evaluation database. The inception of TRECVID (http://trecvid.nist.gov/) has rapidly advanced the research in SBD. As far as by study, it is found that researchers have been able to detect CUTs to some remarkable extent, but the detection of GT still remains a challenging position. For detection of GT, it is really difficult to select an optimal feature set and preparation of cluster which may clearly indicate GT in any type of videos.

The proposed approach for video shot boundary detection can be broadly partitioned into two sub-processes: (1) Model Generation, and (2) Boundary Detection. The proposed work has been compared with (Mohanta, Saha & Chanda, 2012) which is based on Model-Based Shot Boundary Detection Technique Using Frame Transition Parameters. Experimental results showing remarkable success to detect GT as feature set has been selected in such a manner so that proposed model can easily classify boundary and no-boundary shots very efficiently.

Next, briefly describe our proposed approach, which consists of two sub-procedures: (a) Model Generation and (b) Boundary Detection in section Proposed Methodology.