Key Technologies, Challenges and Future Prospects of Neural Network in Automatic Driving System

Abstract

Automatic driving technology is a disruptive force to reshape the future travel and transportation system. Its development has undergone a fundamental transformation from a rule-based modular approach to a data-driven deep learning paradigm. This paper systematically summarizes the key role, frontier progress and core challenges of neural network as the core technology engine in the whole technology stack of automatic driving perception, decision-making, planning and control. Firstly, this paper summarizes the application foundation of convolutional neural network, cyclic neural network, transformer and graph neural network in the field of automatic driving; Then the representative work and implementation path in the core tasks of multi-sensor fusion 3D target detection, interactive behavior prediction, end-to-end driving are analyzed in detail. The analysis shows that although the neural network has greatly improved the performance ceiling of the system, its inherent black box characteristics, vulnerability to adversarial attacks, difficult to deal with the "long tail problem" and high computing costs are still the fundamental obstacles restricting its safe landing. Finally, this paper looks forward to the future research directions of multimodal large-scale model, causal reasoning, neural radiation field simulation and vehicle road coordination, aiming to provide a comprehensive and profound perspective of technology development for researchers in the field.