Enhancing Object Detection Accuracy in Occluded Scenarios Using V2X Cooperative Perception and Deep Learning

Abstract

Occlusion poses a significant challenge to accurate object detection in autonomous driving systems, particularly in dense urban environments where single-agent perception often fails. This research introduces a novel deep learning-based framework, V2X-OccluFusion, designed to enhance detection accuracy by leveraging Vehicle-to-Everything (V2X) cooperative perception and occlusion-aware pre training. The model combines self-supervised masked BEV feature reconstruction with lightweight state-space fusion architecture, enabling multi-agent vehicles to share and reconstruct spatially occluded information. Extensive experiments were conducted across diverse datasets and occlusion levels, including full visibility, partial, and heavy occlusion. Results show that V2X-OccluFusion significantly outperforms baseline models such as Early Fusion and V2X-ViT in both detection accuracy and object recall, especially under heavy occlusion, where it achieved a 25.5% performance improvement over baselines. Additionally, the model demonstrated lower GPU memory usage and faster inference speed, supporting real-time deployment. Communication efficiency was also superior, using less bandwidth while maintaining detection robustness under variable V2X conditions. These findings validate the effectiveness of combining cooperative multi-agent perception with occlusion-aware training for autonomous systems. The research contributes not only to improving detection under occlusion but also sets a foundation for scalable, real-time V2X perception systems adaptable to real-world constraints. The study concludes with recommendations for future enhancements involving multimodal data fusion and federated deployment.