Cooperative Perception for 3D Object Detection in Driving Scenarios Using Infrastructure Sensors

Arnold, Eduardo; Dianati, Mehrdad; de Temple, Robert; Fallah, Saber · 2022 · Crossref

DOI: 10.1109/tits.2020.3028424

archive: archived pipeline: cataloged verified

Get this paper ↗ (DOI — opens at the source; we link to it, we don't host it)

Summary

This paper addresses the limitations of single-point 3D object detection in autonomous driving, specifically regarding occlusion, restricted field-of-view, and low point density at distances. To mitigate these issues, the authors propose a cooperative perception system utilizing infrastructure sensors distributed around the environment. The study evaluates three fusion schemes—early fusion (combining raw point clouds before detection), late fusion (fusing detected bounding boxes after independent processing), and a hybrid approach—using a novel synthetic dataset. The research aims to determine which scheme offers the best balance between detection accuracy and communication bandwidth, while also analyzing the impact of sensor configuration on system performance. The experimental design employs a central fusion system that aggregates data from up to eight fixed infrastructure sensors equipped with depth sensing capabilities (e.g., lidar or depth cameras). The authors generated a bespoke synthetic dataset simulating two complex driving scenarios: a T-junction and a roundabout, chosen for their high occlusion rates. The detection model used is VoxelNet, a deep learning architecture that processes 3D point clouds. In the early fusion scheme, raw point clouds from all sensors are transformed into a global coordinate system and concatenated before being fed into the detection model. In the late fusion scheme, each sensor independently detects objects, and the resulting bounding boxes are fused centrally using Non-Maximum Suppression (NMS) to resolve overlaps. The hybrid scheme combines both by sharing raw data only for distant points (outside a specific radius) and detected boxes for nearby points, thereby reducing communication costs. The results demonstrate that the early fusion scheme significantly outperforms late fusion in detection accuracy. In the most challenging scenario, cooperative perception via early fusion achieved a recall rate of over 95%, whereas single-point sensing achieved only 30%. Late fusion improved upon single-point sensing but remained inferior to early fusion due to the inability to recover objects completely occluded in all individual sensor views. The hybrid scheme offered a middle ground, reducing communication bandwidth compared to early fusion while maintaining higher accuracy than late fusion. The study also found that increasing the number of sensors and optimizing their placement to minimize blind spots significantly enhanced detection performance. The significance of this work lies in demonstrating that infrastructure-based cooperative perception can effectively overcome the inherent limitations of onboard vehicle sensors, particularly in complex urban environments. The findings provide practical insights for deploying such systems, highlighting the trade-offs between detection reliability and communication overhead. By validating the superiority of early fusion for accuracy and the utility of hybrid schemes for bandwidth efficiency, the paper supports the development of safer autonomous driving systems that rely on shared environmental data rather than isolated vehicle perception.

Provenance

The full processing record for this entry. Every stage of this paper's journey through the pipeline is logged — what ran, with which tool and model, how many attempts it took, and when it last completed.

Summary generated by qwen3.6-27b-prismaquant on 2026-06-26; verification: verified.

Topics

Ranked by relevance to this paper. Hover a topic for its definition.

• perceptual countermeasures