High-precision and low-latency object detection with recursive dual-path encoding for event camera

Engineering notes

Extensive experiments conducted on the Gen1 and 1 Mpx automotive event datasets demonstrate that the proposed method achieves state-of-the-art detection accuracy while maintaining ultra-low inference latency. Compared with existing mainstream methods, our approach achieves a superior balance between precision and speed.

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

Event cameras offer sub-millisecond latency, high dynamic range, and asynchronous sparse output, making them a promising bio-inspired sensing paradigm for high-speed visual perception. However, directly applying mainstream object detection frameworks to event data often results in low accuracy, poor adaptability, and an unbalanced trade-off between latency and performance. To address these challenges, this paper proposes an end-to-end, high-precision, ultra-low-latency object detection method tailored for event cameras, achieving an excellent trade-off between detection accuracy and computational efficiency. The proposed method first converts asynchronous event streams into structured spatiotemporal tensors through a tailored preprocessing pipeline, enabling compatibility with standard 2D convolutions. Subsequently, a four-level recursive dual-path progressive interaction encoder is designed to perform hierarchical downsampling. This encoder jointly extracts global and local features via a CNN–Transformer dual-branch architecture, enhanced with a Plain LSTM to stabilize temporal memory integration. To strengthen multi-scale context modeling, a lightweight global context enhancement module that incorporates parallel dilated convolutions and global pooling is introduced, effectively capturing long-range dependencies. Finally, a saliency-guided attention detection head, consisting of separate channel and spatial branches, is employed to emphasize target regions and suppress background interference. Extensive experiments conducted on the Gen1 and 1 Mpx automotive event datasets demonstrate that the proposed method achieves state-of-the-art detection accuracy while maintaining ultra-low inference latency. Compared with existing mainstream methods, our approach achieves a superior balance between precision and speed. This work provides an efficient, high-performance solution for real-time object detection using event cameras in autonomous driving and high-speed robotic vision systems.

Links and sources

Comments