MobileNet V2 SSD Based Pedestrian Detection System Using Normalized 360° Image

Abstract

The growth of vehicular traffic in various areas such as urban and rural areas is getting
higher due to the increasing need for transportation equipment. This condition causes many scientists to continuously improve the quality of the mode of transportation so that it is easier, safer, and more practical to use. As a result, the idea of driverless transportation has emerged to meet this need. The form of transportation is private vehicles, and it is expected to be in the form of mass vehicles such as buses or trains. Several aspects must be considered in designing autonomous transportation or autonomous electric vehicles so as not to cause accidents that can endanger the driver and the surrounding environment. Among them is the existence of a pedestrian detection system. This system is critical because, like conventional vehicles, autonomous vehicles must avoid pedestrians, but they work with no driver assistance. In this study, a software system that can detect pedestrians from all directions using a 360° camera was developed to overcome the above. This system also utilizes deep learning technology. The design and realization of this system went through several stages. The stages include installing supporting software on the NVIDIA Jetson AGX Xavier, capturing video data with a 360° camera for composing a dataset of 19,038 images, training the MobileNet V2 SSD with the dataset, and testing the trained model with the real-time and offline testing process. As a result, by testing 548 normalized 360° images offline for daytime conditions, 60.40% of images were perfectly detectable, while for 514 normalized 360° images for evening conditions, 62.25% of images were ideally detected.