计算机视觉
计算机科学
人工智能
变压器
目标检测
事件(粒子物理)
工程类
电压
电气工程
分割
物理
量子力学
作者
Heng Zhang,Nuo Chen,Miao Li,Wei An
标识
DOI:10.1109/isctech63666.2024.10845340
摘要
Unmanned aerial vehicle (UAV) detection tasks have always been a hot issue in the field of computer vision. Event cameras focus solely on the motion within a scene, featuring high dynamic range and low latency. For small UAVs with high speeds, event cameras can acutely perceive their motion information in various complex scenarios. Therefore, in this work, we utilize event cameras to capture event stream motion data of UAVs with low background redundancy. Simultaneously, we combine the Swin Transformer which pays attention to local spatial correlation with the Spiking Neural Network (SNN) which is sensitive to temporal features, and propose an energy-efficient backbone network for spatiotemporal feature extraction of small targets—Spiking Swin Transformer (SST). Utilizing the latest advancements in spike backpropagation research—surrogate gradient learning and Spikingjelly framework, based on Feature Pyramid Networks (FPN) and Single Shot MultiBox Detector (SSD), we propose a motion small target detection algorithm that can be directly trained on event stream datasets with SST as the backbone network. In this paper, we conducted experiments on our self-made event stream UAV object detection dataset and established a new state-of-the-art SNNs detection result.
科研通智能强力驱动
Strongly Powered by AbleSci AI