Estimating 3D Coordinates of Bounding Box Center after Detection the Object in Image of Cluttered Environment Based on Stereo Vision and U-Net for navigate the robots

• Published in: 2020 28th Iranian Conference on Electrical Engineering (ICEE) pp. 1 - 6
• Main Authors: Dashti, Alireza Nasirzadeh, Nooshyar, Mehdi, Akbarimajd, Adel
• Format: Conference Proceeding
• Language: English
• Published: IEEE 04.08.2020
• Subjects: camera calibration; Cameras; computer vision; Decoding; deep convolutional network; Image segmentation; Object detection; Stereo vision; Three-dimensional displays; Training; u-net
• ISSN: 2642-9527
• DOI: 10.1109/ICEE50131.2020.9260689

Recently deep learning attained to good achievements in various fields and stereo vision is also one of the most important methods for estimating 3D coordinates. In this paper, two methods of deep learning and stereo vision are combined and an end to end method is presented that also does the object detection easily with high accuracy and estimates 3D coordinates of bounding box center that surrounds the observable part of object in cluttered environment image. Proposed network based on U-Net. Classical U-Net architectures is much famous in segmentation of satellite and medical images, but we made some modifications in it that convert the u-net to a detector and mask generator for object in image. In addition, this method can act as a 3D coordinates estimator by reusing the u-net decoder part as a feature extractor and add some fully connected layers to it. proposed network can obtain acceptable results with shorter training times in a small data set and it can useful for navigate the robots. we created a dataset containing 500 pairs of images from 10 objects that are cluttered in each image. the model was evaluated on mentioned dataset and The results was impressive.