Instance segmentation for roadside objects using a simulation environment

• Main Author: Wang, Sijie
• Other Authors: Tay, Wee Peng
• Format: Thesis-Master by Coursework
• Language: English
• Published: Nanyang Technological University 2021
• Subjects: Engineering::Computer science and engineering::Computing methodologies::Image processing and computer vision; Engineering::Electrical and electronic engineering
• Online Access: https://hdl.handle.net/10356/149566
• Institution: Nanyang Technological University

In the autonomous driving system, the understanding of traffic is always an important task. Especially in the field of the detection and recognition for roadside objects, it can help to guide vehicles, prevent them from deviating, and assist them in positioning and localization. To achieve the goal, in these years, deep learning and computer vision technology have been powerful tools for instance segmentation for roadside objects. In addition, with the continuous advancement of computer technology and hardware, it has been possible to train and test instance segmentation algorithms in autonomous driving simulation environments. Compared with collecting data in real environment, the simulation environment can directly generate data through computing, which saves a lot of manpower, time and financial resources. In this dissertation, a method for generating instance segmentation labels using point clouds and semantic labels is proposed, and the instance segmentation algorithm, Mask R-CNN, is evaluated on the dataset generated from CARLA simulator. The final result shows that Mask R-CNN on CARLA has achieved the best performance compared with other baselines.