Intermediate Feature Fusion Method for Collaborative Perception of Vehicles and Infrastructure

The single vehicle intelligent system has inherent limitations in the perception level, such as insufficient over the horizon capability, lack of global perspective and limited stability of extreme scenes. The vehicle-to-infrastructure group intelligent collaborative perception technology provides an effective solution to the above problems through information sharing and fusion among multi-agent. Between data level fusion and decision level fusion, intermediate feature fusion achieves the balance between communication efficiency and perception accuracy by realizing the interaction and integration of multi-source information at the feature level, which has become a research hotspot in the field of vehicle-to-infrastructure collaborative perception. This paper reviews the development process of intermediate feature fusion, and reveals the evolution logic of collaborative perception from "result sharing" to "feature collaboration" from the paradigm shift from late collaboration to medium-term collaboration; The three core technologies of feature extraction and representation, spatial feature alignment, feature compression and communication optimization are analyzed; Four typical paradigms of global fusion, sparse fusion, hierarchical fusion and attention fusion and their application scenarios are summarized; The research hotspots of transformer architecture, task oriented communication optimization and the key challenges faced by system deployment are discussed. On this basis, the future directions of generative AI empowerment, evolution from collaborative reasoning to collaborative generation, and standardized deployment are prospected, in order to provide reference for the in-depth research and engineering implementation of intermediate feature fusion technology.