Object classification in urban environments for autonomous navigation

Abstract

Depending on the environment, urban, office or home, different objects can be identified in the scene during machine-human interaction. An essential task to facilitate object recognition in natural scenes is image segmentation, due to segmented regions can be analyzed separately. In this work, we propose the classification of objects in urban context images using image segmentation and convolutional neural networks (CNNs). A training set was created by extracting patches in color images of the Kitti-360 database, operating as a reference the semantic images for only encircling the object of interest. Eight categories of everyday objects found commonly in urban places were selected. Different CNN models were trained using our gathered set. During experimental tests, a segmentation approach based on CIELab and complex-space was used to provide the most representative regions of the image. Selected bounding boxes are classified per image instead of the whole image. Our classification results show high-accuracy values using images of small size but strategically selected based on the segmented image.