EventEye: Towards High-Frequency Perception Enhancement for Autonomous Vehicles Using Infrastructure-Mounted Event Cameras

Engineering notes

EventEye achieves an error of around 0.3 m with a processing latency of less than 25 ms.

Chinese explanation / 中文解读

中文解读待补充：本站会优先为端到端自动驾驶、BEV感知、3D目标检测、轨迹预测、路径规划、LiDAR感知等高价值论文补充中文说明。

Original abstract

Infrastructure-assisted autonomous driving is an emerging paradigm expected to greatly enhance the safety of autonomous vehicles. Current systems primarily use LiDAR or RGB cameras as the infrastructure sensors, which suffer from high end-to-end latency and environmental susceptibility, respectively. In this work, we propose EventEye, the first system that integrates event cameras into infrastructure to enhance a vehicle’s real-time perception by seamlessly fusing high-frequency event data with the vehicle’s LiDAR data. The core concept of EventEye is to leverage the inherent sparsity of event camera data on the infrastructure to estimate Bird’s-Eye-View (BEV) occupancy maps with grid flows recurrently. The generated occupancy map is then merged in real time with vehicle LiDAR-generated maps. We implement EventEye on a real-world testbed and evaluate the performance of EventEye across various scenarios. EventEye achieves an error of around 0.3 m with a processing latency of less than 25 ms. Compared to previous systems, this enables 51% earlier detection of incoming objects, providing autonomous vehicles with crucial additional reaction distance.

Links and sources

• Official / arXiv page

Comments