CollabGAT: Collaborative Perception Using Graph Attention Network

Abstract

Collaborative perception exploits exchanging perception data across multiple agents to enhance situational awareness. To overcome the constraints created by limited network resources, it is necessary to have a reliable collaborative perception that can sustain the trade-off between performance and bandwidth. This research proposes a method to address this issue through the development of a graph attention network (GAT) for intermediate collaborative perception. The proposed approach aims to enhance object detection accuracy by exchanging information among multiple agents, addressing issues such as sensor limitations, and occluded objects, and expanding the sensing range. Additionally, the proposed approach addresses the challenges caused by limited communication bandwidth and inconsistencies in the data shared by different agents by maintaining a balance between performance and bandwidth, fitting real-life application requirements. Our approach aggregates the intermediate features obtained from multiple neighboring agents and introduces a novel attention mechanism to selectively emphasize significant regions within the intermediate feature maps. This attention mechanism operates on both the channel and spatial levels to direct the data aggregation process. This research presents a new method for aggregating features utilizing various attention architectures. We perform quantitative and qualitative assessments of the final results of this method and compare them to other state-of-the-art collaborative perception techniques. The results of this work are evaluated using the V2XSim and OPV2V datasets, and our quantitative and qualitative experiments in multi-agent object detection show that our approach achieves a better performance than the state-of-the-art collaborative perception methods.