How Emergent Architectures Solve the Bird’s Eyes View Shortcomings: A Systematic Review

Abstract

Autonomous driving requires robust, accurate and real-time environmental perception. Although cameras, LiDAR, and radar provide complementary sensing capabilities, individual modalities remain vulnerable to limitations such as depth ambiguity, point-cloud sparsity, adverse weather, and restricted angular resolution. Multi-sensor fusion therefore represents a key approach for improving 3D object detection and scene understanding. This systematic review analyzes research published from 2022 to 2026 using the PRISMA framework and studies retrieved from Scopus and Web of Science. The review examines sensor characteristics and major fusion strategies, including early, feature, late, and hybrid fusion. It also covers BEV representations, Transformer-based architectures, evaluation methods, and benchmark datasets. Particular attention is given to multimodal combinations involving cameras, LiDAR, and radar, as well as cooperative perception through V2X systems. The analysis shows a clear shift toward BEV-based feature fusion and attention-driven architectures, reflecting their ability to integrate heterogeneous spatial and semantic information. However, the literature remains fragmented in several critical areas. Real-world benchmarks covering adverse weather and challenging conditions are still limited, while computational requirements remain a major barrier to real-time onboard deployment. Robust online calibration, adversarial security, cross-domain generalization, and heterogeneous cross-manufacturer V2X perception are also insufficiently investigated. In addition, 3D occupancy representations remain less explored than conventional bounding-box detection, and emerging architectures such as Mamba and large language models require broader validation on real-world data. These findings reveal persistent gaps between research advances and practical deployment, highlighting the need for efficient, secure, weather-robust, and generalizable fusion approaches for heterogeneous sensors and cooperative autonomous driving scenarios in real-world applications.