Cooperative Perception with Deep Reinforcement Learning for Connected Vehicles

Aoki, Shunsuke; Higuchi, Takamasa; Altintas, Onur · 2020 · Crossref

DOI: 10.1109/iv47402.2020.9304570

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 individual vehicle perception systems, which suffer from restricted coverage and detection accuracy due to occlusions by other objects. While cooperative perception—where connected vehicles exchange sensor data via Vehicle-to-Vehicle (V2V) communications—can mitigate these blind spots, it generates significant network traffic. In congested environments, redundant data transmission leads to packet collisions and loss, undermining communication reliability and safety. The authors propose a deep reinforcement learning (DRL) approach to solve the "Message Selection Problem," enabling vehicles to intelligently select which perception data to transmit, thereby reducing network load while maintaining or enhancing detection accuracy. To implement and evaluate this scheme, the authors developed the Cooperative & Intelligent Vehicle Simulation (CIVS) Platform. This framework integrates SUMO for traffic simulation, CARLA for realistic vehicle and sensor simulation, and a YOLO-based object classifier. The DRL model uses a Deep Q-Network with Convolutional Neural Networks (CNNs) to determine transmission actions. The state input consists of a grid-based circular projection of the vehicle’s surroundings and the current network congestion level. The action space is binary: transmit or discard a Cooperative Perception Message (CPM). The reward function is designed to maximize the detection of objects not visible to the receiver while penalizing redundant information sharing and high network congestion. The evaluation compared the proposed DRL scheme against a baseline protocol where vehicles broadcast CPMs whenever objects are detected. Experiments were conducted across various vehicle densities (50 to 200 vehicles) using two distinct map scenarios for training and testing. Results indicate that the DRL approach significantly improves communication efficiency. Specifically, the average packet reception ratio increased by up to 27% compared to the baseline, as the intelligent selection of data reduced network congestion and packet collisions. Furthermore, despite transmitting less data, the scheme enhanced the average object detection ratio by up to 12%. This improvement is attributed to the reduced packet loss, which ensured that critical perception data reached neighboring vehicles reliably. The study demonstrates that deep reinforcement learning can effectively balance the trade-off between network resource usage and perception accuracy in cooperative driving systems. By mitigating redundant transmissions, the proposed method enhances the reliability of V2V communications, which is critical for road safety. The development of the CIVS platform also provides a reusable, integrated simulation environment for testing cooperative perception algorithms, addressing the challenges of testing such systems with real vehicles.

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.

• v2x connected vehicle

Information type

What kind of knowledge this paper contributes, grouped by family — independent of topic (what it is about) and method (how it was studied).

• Methodological Resource: tool software