On the Problem of Restoring and Classifying a 3D Object in Creating a Simulator of a Realistic Urban Environment

Area of Science:

• Computer Vision
• Artificial Intelligence
• Robotics

Background:

• The rapid motorization since the 20th century necessitates enhanced vehicle safety and reduced environmental impact.
• Managing traffic flow is challenging due to unpredictable human driver behavior.
• Current training methods for unmanned vehicles involve extensive real-world driving, which is time-consuming and cannot cover all scenarios.

Purpose of the Study:

• To develop an intelligent vehicle recognition system for creating optimized 3D mesh objects for realistic urban simulators.
• To reduce the training time and computational costs associated with developing unmanned transport systems.
• To enable the generation of diverse road scenarios within a simulator.

Main Methods:

• Development of a method for recreating realistic environments using monocular cameras.
• Application of convolutional neural networks (CNNs) for intelligent vehicle recognition and 3D object mesh generation.
• Modification and development of a novel CNN architecture to balance accuracy, object optimization, and computational complexity.

Main Results:

• A novel CNN architecture was designed and modified, enabling accurate object classification.
• Optimized mesh objects were generated, suitable for use in a simulator without system overload.
• Reduced computational complexity was achieved compared to existing neural network architectures.

Conclusions:

• The proposed intelligent vehicle recognition system effectively creates optimized 3D mesh objects for realistic urban simulators.
• The developed methods significantly decrease training time and computational expenses for unmanned transport systems.
• This approach facilitates more efficient and comprehensive training of autonomous vehicles.