Multi-Task Collaborative Perception Algorithm Based on Consensus Perception

This Multi-agent collaborative perception is a hot topic in the current field of machine learning. Through communication and sharing of perceived scene information between agents, this approach can tackle issues such as occlusion and sparse observations present in single-agent observations. Existing studies have shown that localization errors are a significant factor affecting the accuracy of perception results. Furthermore, most prior work has focused solely on the performance of algorithms in individual tasks without considering multi-task scenarios. In this paper, a multitask collaborative perception method based on consensus perception is proposed. It integrates local perception information from each intelligent agent to form a global perception consensus, which is then used to correct the spatial feature maps of the agents, alleviating the impact of localization errors on perception results. Additionally, the use of a feature selection module based on spatial confidence maps and a multiscale attention feature fusion module optimizes the performance of the collaborative perception process. By setting up multiple decoders in parallel, a single model is able to handle diverse tasks. Experimental results demonstrate that the proposed consensus-based collaborative perception method effectively mitigates the negative impact of localization errors on perception performance, and the algorithm can achieve good perception results across multiple tasks.