Indian Journal of Science and Technology

Abstract

Objectives: This paper proposes a technique of segmenting a target object more robustly by combining and clustering 2-dimensional and 3-dimensional features of 3D stereoscopic images coming sequentially. Methods/Statistical Analysis: The proposed object detection technique first applies a disparity computation algorithm to the left and right stereo input images that are captured to extract the depth values representing the distance between a camera and a target object by each image pixel. The technique then effectively clusters color and depth features using feature similarity. Subsequently, the method excludes the background area from the image and finally detects an object of interest. Findings: The experimental results of this paper show that the proposed 3-dimensional feature-based target detection technique extracted objects more robustly than other existing object detection methods. To evaluate the performance of the proposed algorithm of detecting an object of interest, this paper uses various types of indoor and outdoor input images without any particular constraints. To compare the performance of the suggested object segmentation method, this paper defined root mean square error measure. The measure which is the scale to deal with general and particular viewpoints of image quality is often used to measure the difference between an actual value and a measured value. It is known as a good scale tool for accuracy measurement. The two existing methods use 2D features only to segment an object of interest, and thereby include many false positive errors, whereas the proposed method effectively clusters the 3-dimensional distance feature as well as 2-dimensional feature so that it segments an object of interest more accurately than the other two methods in terms of quantitative aspect. Improvements/Applications: It is expected that the proposed technique of detecting an object of interest will be used in various types of actual application areas related to multimedia contents.
Keywords: Clustering, Color model, Feature, Multimedia content, Objects of Interest, Segmentation