A Real-Time for Classification of Moving Objects

Introduction  Developed a software system by a static to detect and track moving object  Main components: 1.initialization 2.adaptive update of a background model 3.detection and tracking of moving objects 4.extraction of feature vectors and classification

Change Detection and Background Modelling  Temporal differencing -adaptive to changes in the environment, but does not detect the entire object  Background subtraction - provide more reliable information about moving object but requires more complex processing

Background Initialization  Background initialization is done in the first 1-2 seconds.  First, initialize the background image by the first frame -  Create a binary mask - use the Eucledian metric for measuring the distance

 Looking at the binary mask created by thresholding the difference between and  Updating the background pixels as:

 The process stops when the number of remaining suspicious pixels  After the process is finished the uninitialized background pixel set their values from the last frame

Background Adaptation  An adaptive update of the background due to the two main reasons -changes caused by moving objects -changes due to illumination

Target Detection and Tracking  Target detection is performing using the background subtraction.  The target is partitioned into horizontal strips and the color table entries hold average RGB values for each strip  The number of strips depends on target size

 The color table is used for defineing an individual threshold for each pixel and every target.  is the threshold image  is the target color table of size

Feature Vectors and Classification  The ratio between perimeter and area is commonly used as the object shape characteristic  The ratio between the lengths of the vertical and horizontal axes of the ellipse fitted to the object contour

 Star skeleton is created by connecting the center of mass of the moving object with contour points corresponding to the local maxima of the distance

 Classification is performed for every 30 consecutive frames.