Analysis and Comparison of Event Stream Coding Algorithms for Object Detection Based on Spiking Neural Networks

Since the introduction of event cameras, the way these biomimetic devices process visual information has garnered significant attention from researchers. The processing of event camera data has been deeply concerned by researchers due to the manner in which these devices handle visual information. Event cameras have unparalleled advantages in object detection tasks implemented with deep learning methods compared to traditional frame cameras, such as higher frame rate performance under the same hardware resource consumption. In this paper, we summarize the commonly used event stream encoding methods, detailing their pros and cons. Subsequently, we propose an end-to-end event stream encoding learning module for object detection, which replaces the manual encoding module, integrates with the object detection network, e.g., Vgg, Mobilenet and Densenet, and is co-trained with it. With only a 1k increase in the overall network model size, the mAP, a metric describing the network performance, is improved by up to 14%, achieving energy-saving embedded applications.