摘要
针对传统相机在拍摄人眼运动时易产生动态模糊、时间分辨率低等问题,采用事件相机近眼拍摄构建Spiking-Eye数据集,并提出一种坐标注意力的脉冲神经网络模型(CA-SpikingRepVGG)。模型读取编码后的事件数据,经过带坐标注意力的主干网络进行特征提取,最后馈入检测头进行检测。实验结果显示:CA-SpikingRepVGG的平均检测精确率R_(P)达到了70.8%,与SpikingVGG-16比较,该模型的R_(P)提高了15.9%,召回率R_(r)提高了14.2%;仅需SpikingDensenet模型1/3的训练时间,比其R_(P)提高1.8%、R_(r)提高0.9%。结果表明:该模型在针对眼球运动这一场景下对人眼的检测追踪能力更强,可以很好地完成注视估计任务。
The problems of dynamic blur and low temporal resolution in capturing eye movements with traditional cameras are addressed by employing an event camera for close-range capture and constructing a spiking-eye dataset.A spiking neural network model with a coordinate attention referred to as CA-SpikingRepVGG.The model reads encoded event data and performs feature extraction using the attention-based backbone network,followed by detection using the detection head.Experimental results demonstrate that CA-SpikingRepVGG achieves a mean average precision R_(P)of 70.8%.Compared to SpikingVGG-16,the model shows a 15.9%improvement in R_(P)and a 14.2%increase in R_(r).With only one-third of the training time required by SpikingDensenet,the model achieves a 1.8%improvement in R_(P)and a 0.9%improvement in R_(r).These results indicate that the proposed model exhibits stronger eye detection and tracking capabilities in the context of eye movement,effectively accomplishing gaze estimation tasks.
作者
王红霞
赵志国
WANG Hongxia;ZHAO Zhiguo(Shenyang Polytechnic University,Shenyang,Liaoning 110158,China)
出处
《计量学报》
CSCD
北大核心
2024年第7期982-988,共7页
Acta Metrologica Sinica
基金
辽宁省自然科学基金(2022-MS-276)。
关键词
机器视觉
目标检测
脉冲神经网络
注视估计
坐标注意力
召回率
事件相机
machine vision
object detection
spiking neural network
gaze estimation
coordinate attention
recall
event camera