Article Preview
TopI. Introduction
Tracking and assessment of the activity in a moving body is a challenging task. The evaluation of object in motion and distinguish the background and foreground, many techniques have been implemented. This includes optical flow algorithm, interframe difference and background subtraction(Kim & Jung, 2017). The optical flow algorithm proposed by researchers (Fuentes et al., 2018) requires estimation of flow vectors computing the magnitude of flow estimations and warping flow fields. But the technique cannot deal with dynamic conditions. The temporal frame differencing as per researchers (Chiu et al., 2018) exploit the consecutive frames to mine the pixel by pixel difference of the two images. The subtraction process cannot handle the fast moving objects because when the object in motion goes still or moves fast, for a few frames, background and foreground can’t be distinguished. Also, thresholding the difference is a kind of exacting approach which may lead to unnoticed activity of the interest of researcher (Anandhalli & Baligar, 2015)(Mahajan & Padha, 2018). The widespread modus operandi for distinguishing the background from the foreground is Background Subtraction (BS) which extracts low entropy based pixels of the object in motion without any prior information about the scene proving to be effectual in stationary camera arrangements and highly precise in pixel, frame as well as region level procedure (Kumar & Yadav, 2016b). The procedure of background subtraction comprise of three stages: a) initialize the background by mapping the spatio-temporal constraints and reconstruct the model based on intensity variations, b) Foreground extraction which detects and segments the moving object and c) Maintenance of background by updating of pixel variance in spatial domain along temporal constraints in the video processing of the sequences.
The techniques implemented by various researchers (Kumar & Yadav, 2016a) (Fazli et al., 2009a) (Fazli et al., 2009b) in these three stages define the accuracy and precision of the proposed algorithms. In our study, in the background modeling stage, an improved Expectation Maximization based algorithm for Gaussian mixtures is proposed to initialize the multi modal background with dynamic conditions. In the foreground detection stage, a curvelet based denoising method with adaptive normalization of the intensities values is done to minimize the intraclass variance of the pixels for the removal of isolated background pixels and filling of the isolated foreground pixels. In the background maintenance stage, the updation of pixel variance in spatial domain is done by removing all connected components (objects) that have fewer than P pixels, producing another image sequence and the convolution of image vectors in done in temporal domain. This is illustrated in Fig. 1.
Figure 1. Tracking of the foreground moving object
Section II contains the relevant research based on the literature surveyed, Section III contains the proposed technique for the construction and updating of the background model, Section IV contains the Experimental results and Section V contains the conclusion and future scope.