PerceptNet-V2X: Perception Network for Vehicle to Everything Scenarios in Autonomous Driving

Engineering notes

The framework integrates the FPV representation with state-of-the-art detectors such as YOLOV12 and RT-DETR, demonstrating flexible and modular deployment. On the OPV2V synthetic benchmark, PerceptNet-V2X achieves 95.63% AP@0.5 and 94.36% F1@0.5, representing improvements of up to 5.02% AP and 3.15% F1 over traditional BEV representations.

Chinese explanation / 中文解读

中文解读待补充：本站会优先为端到端自动驾驶、BEV感知、3D目标检测、轨迹预测、路径规划、LiDAR感知等高价值论文补充中文说明。

Original abstract

Collaborative perception is essential for autonomous vehicles (AVs) to overcome individual sensing limitations in occluded or complex traffic environments. This paper introduces PerceptNet-V2X, a novel intermediate collaboration framework for scalable perception in Vehicle-to-Everything (V2X) settings. At its core is the Full Perspective View (FPV), a new compact 2D representation of 3D point clouds that preserves critical geometric structure while enabling effective multi-agent spatial reasoning. Unlike conventional Bird’s Eye View (BEV) projections, FPV mitigates data sparsity and enhances compatibility with modern 2D detection architectures. The framework integrates the FPV representation with state-of-the-art detectors such as YOLOV12 and RT-DETR, demonstrating flexible and modular deployment. Comprehensive evaluations on the OPV2V and V2X-Real datasets validate FPV’s impact on collaborative perception performance. On the OPV2V synthetic benchmark, PerceptNet-V2X achieves 95.63% AP@0.5 and 94.36% F1@0.5, representing improvements of up to 5.02% AP and 3.15% F1 over traditional BEV representations. On the V2X-Real real-world dataset, the framework delivers 61.55% mAP@0.3, a 16% gain over prior approaches. Overall, this work contributes a scalable, modular and effective collaborative perception solution centered on the FPV representation, advancing the reliability of multi-agent autonomous systems. Code available at: https://github.com/brumocas/PerceptNet_FuseNet_V2X

Links and sources

• Official / arXiv page

Comments