Region and Decision Tree-based segmentations for Multi-objects detection and classification in Outdoor Scenes

Abstract

Accurate segmentation and detection of all mixed and occluded objects in the complex indoor/outdoor scenes become a vital topic of computer vision. These above-mentioned scenarios are the significant part of important vision applications such as autonomous driving, traffic monitoring, security surveillance, humane body parts detection, objects tracking and scene recognition. It is still difficult to accurately detect all the objects in the image due to illumination changes, occlusion and different directions. Meanwhile, segmenting the image in parts helps to detect the multi objects accurately. In this paper, we designed a system having improved techniques for the accurate segmentation and detection of multi objects. Firstly, we have combined the results of two methods for accurate segmentation of multiple objects, (i) Decision trees for labeling every neighboring pixel and assigning a separate color to all object present in complex images and (ii) Region-based segmentation for significant detection of multiple regions and drawing boundaries of all objects. Secondly, detection is performed by searching the objects with previously assigned colors. Finally, we have performed labeling with class name to all objects present in the images. We have performed our experimental over two benchmarked datasets as Instance Saliency images and MSRC. Our experimental work has shown improved results with respect to other state of the art algorithms.