Collaborative 3D Object Detection by Intelligent Vehicles Considering Semantic Information and Agent Heterogeneity
Collaborative perception can significantly enhance perception performance through information sharing among multiple smart vehicles and roadside perception systems. Existing methods often require sharing and aggregating all information from other collaborators.
However, in practice, most information shared by collaborators is largely irrelevant to the final perception, making it challenging to identify distant and occluded objects. To address this issue, we propose a novel collaborative framework, named Semantic Aware Heterogeneous Network (SAHNet), which shares and fuses perceptually crucial and useful information among heterogeneous collaborators to improve the performance of 3D object detection. Specifically, we firstly design Foreground and Boundary Feature Selection (FBFS) to enhance meaningful feature extraction. A Heterogeneous Feature Transfer module (HFF) is then proposed to consider collaborators' heterogeneity to better transfer perception-critical features. Finally, we put forward a Semantic Feature Fusion module (SFF) that effectively aggregates features using semantic information.
METHOD
- The essential problem in intermediate fusion-based collaborative perception is to allow multiple collaborators to share perceived features via communication networks to promote each agent's perception capability.
However, the majority of the features perceived by each agent are not relevant for object detection. For communication efficiency and object detection accuracy, we focus on maximizing each agent's perception capacity by selecting, transferring, and fusing the most important information.
-
For the 3D object detection in the collaborative perception, we need to focus on valuable features. In such a process, it is crucial to consider both foreground and boundary features because the literature has shown that they are essential to object classification and localization. To address this issue, we propose a lightweight foreground boundary attention module.
-
In a collaborative perception system, each collaborator has its own distinct characteristics. Their sensors, for example, will be affected by factors such as hardware quality and acquisition parameters, impacting the quality of the output 3D point cloud. Besides, there are generally two different types of agents in collaborative perception system: infrastructure (road-side sensing system) and autonomous vehicles (AV). Each of them can share perceptive information with other agents, providing multiple viewpoints for the same scene. However, it is important to note that an infrastructure's LiDAR installations are positioned at elevated locations, having a broader field of view with minimal occlusions, while onboard LiDAR can provide more detailed information about the surrounding environment. In order to encode this kind of heterogeneity during the transmission process, we propose a heterogeneous feature transfer module (HFF) that incorporates the types of collaborators into consideration.
RESULTS
We evaluate the effectiveness of the proposed AEV2V on the V2V perception dataset V2XSet.
Visualization of collaboration
CONCLUSION
In this paper, we have developed a novel deep learning based framework for collaborative object detection to improve the perception capabilities of autonomous vehicles. It consists of multiple modules designed to learn critical information from collaborating vehicles and address data heterogeneity. This framework can integrate the learned critical features into the final information fusion, assisting single vehicle in being informed about distant occluded objects and enhancing decision-making capabilities. Extensive experiments conducted on three public datasets demonstrate the effectiveness and robustness of the proposed method, showcasing better performance than the state-of-the-art methods. We believe that our research results contribute to improving the accuracy of collaborative perception in autonomous driving. Future work will investigate methods to compensate for delay in feature transmission and adjust object pose and continue to deepen our understanding of collaborative 3D object detection.