摘要
事件相机是一种启发式传感器,它通过感知光线强度变化输出事件,响应异步和稀疏事件形式的像素级亮度变化,缓解了传统相机在光线条件变化复杂和物体高速运动场景下成像不清晰的问题。最近,基于学习的模式识别方法将事件相机的输出转化为伪图像的表示形式,在光流估计、目标识别等视觉任务中取得了巨大的进步。但是,这类方法丢弃了事件流之间的时间相关性,导致伪图像的纹理不够清晰,特征提取困难。为此,提出了基于事件流划分算法的神经网络框架,显式地融合了事件流的时间信息。该框架将输入的事件流划分成多份,使用权重分配网络给每一份事件流赋予不同的权重,并使其通过卷积神经网络融合时空信息、提取高级特征,最后对输入分类。在N-Caltech101和N-Cars数据集上进行的对比实验表明,与现有最先进算法相比,所提框架在分类准确率上有明显的提升。
Event cameras are asynchronous sensors that operate in a completely different way from traditional cameras.Rather than catching pictures at a steady rate,event cameras measure light changes(called events)separately for every pixel.As a sequence,it alleviates the problems of traditional cameras in complex light conditions and scenes where objects move at high speed.With the development of convolutional neural networks,learning-based pattern recognition methods have made great progress in visual tasks such as optical flow estimation and target recognition by converting the output of the event camera into a pseudo-ima-ge representation.However,such methods abandon the temporal correlation between the event streams,so that the texture of the pseudo image is not clear enough,and it is difficult to extract the features.The key to solving this problem lies in how to model relevant information between events in the sample.Therefore,a neural network framework based on event stream partition algorithm is proposed,which explicitly integrates the temporal information of event streams.The framework divides the incoming stream of events into several parts,and a weight distribution network assigns different weights to each piece of the streams.Then,the framework uses convolutional neural network to fuse temporal information and extract advanced features.Finally,the input sample is classified.We thoroughly validate the proposed framework on object recognition.Comparison experiments on N-Caltech101 and N-cars datasets show that the proposed framework has a significant improvement in classification accuracy compared with the most advanced existing algorithms.
作者
徐化池
史殿习
崔玉宁
景罗希
刘聪
XU Hua-chi;SHI Dian-xi;CUI Yu-ning;JING Luo-xi;LIU Cong(National Innovation Institute of Defense Technology,Beijing 100071,China;College of Computer,National University of Defense Technology,Changsha 410073,China;Tianjin Artificial Intelligence Innovation Center,Tianjin 300457,China)
出处
《计算机科学》
CSCD
北大核心
2022年第5期43-49,共7页
Computer Science
基金
国家重点研发计划(2017YFB1001901)
天津市智能制造专项资金项目(20181108)。
关键词
时间信息
事件流
融合
权重分配
卷积神经网络
Temporal information
Event streams
Fusion
Weight allocation
Convolutional neural network
