Integrated Use of Neural Networks of Various Types in the Task of Detecting Obstacles on the Path of an Unmanned Ground Vehicle

The paper considers the issue of the feasibility and necessity of developing an optimized method for determining static and dynamic obstacles of various classes that arise along the path of an unmanned ground vehicle. It is proposed to use two different types of neural networks as the main tool for solving the problem: a convolutional neural network for the initial identification of potential obstacles and a deep learning neural network—a multilayer perceptron used to analyze and filter potential obstacles based on their position in the frame and the position of the unmanned ground vehicle, traffic lanes, and the location of other elements on the image. The described method identifies not only dynamic but also static obstacles, registers their classes and provides the ability to save information about the object’s class for further target tracking and assessment of the possible threat to all road users including vehicles, people, etc.