Feature Fusion Methods for Multimodal Image Recognition Based on Autonomous Driving

Abstract

With the widespread use of automatic driving technology, multimodal fusion technology has gradually formed three mainstream fusion levels: early stage, middle stage, and late stage. On the basis of extensive literature research, this paper not only systematically analyzes the research progress and optimization direction of multimodal fusion technology from the perspective of feature fusion, but also analyzes the strategies of different feature fusion based on different sensors. What's more, the general data set of multimodal fusion is integrated, and the applicable scenarios of different methods are compared and analyzed. For Camera-LiDAR feature fusion, it is more suitable for urban target detection. Camera-Radar feature fusion completes the task of bad weather perception. BEV multimodal feature fusion uses a unified bird 's-eye view as a shared representation space to realize panoramic perception. At last, the paper summarizes the core problems faced by multimodal fusion technology. In the future, it is still necessary to make continuous breakthroughs in fusion architecture innovation, robustness optimization, and engineering implementation, so as to promote the large-scale application of high-order autonomous driving technology.