Software & Systems

Articles of journal № 4 at 2020 year.

Order result by: Public date | Title | Authors

21. Experimental analysis of the accuracy and performance of varieties of YOLO architectures for computer vision problems [№4 за 2020 год]
Authors: P.A. Bokov , P.D. Kravchenya
Visitors: 3698
Unmanned vehicles are increasingly being introduced into everyday life. To achieve full autonomy when traveling, unmanned vehicles use computer vision systems, which are responsible for analyzing the status of traffic lights, signs, and other objects that can appear on the road. Today, the standard in this area is YOLOv1 architecture, however, it is already obsolete. In this regard, a computer vision sys-tem for an unmanned vehicle based on modern technologies is being developed. There is the problem of choosing a computer vision architecture that will be responsible for analyz-ing traffic. First of all, it must be fast and accurate, because road traffic changes very quickly, and the accuracy of determination directly affects the degree of involvement of passenger-drivers in the pro-cess in order to avoid emergency situations. In addition, the architecture should occupy as little compu-ting power as possible, and not waste a large number of energy resources. To investigate these issues, it was decided to carry out an experiment that would reveal the advantages and disadvantages of various YOLO architectures. Also, the data provided by different researchers is very different due to using dif-ferent equipment while training and testing of networks. This makes it impossible for data to be com-pared objectively. The paper analyzes various types of YOLOv3 architecture and its versions for low-power compu-ting systems YOLOv3-tiny, describes their advantages and disadvantages for computer vision systems. The experiments are carried out on single hardware for all analyzed architectures. Experimental re-search on the accuracy and performance of various YOLO architectures is being done. The VOC2012 dataset is used for training and testing. As a result of the research, the strengths and weaknesses of the architectures under consideration are determined and options for the further development of the tech-nology are analyzed, taking into account the growth in the power of computing systems and the emer-gence of new technological solutions.

◄ ← Preview | 1 | 2 | 3