时域感兴趣区域精确定位与膜电位多核调整的 动态视觉传感器数据分类

Dynamic vision sensors data classification based on precise temporal region of interest locating and multi-kernel membrane potential adjusting

中文关键词  动态视觉传感器DVS  DVS数据分类  目标识别  时域感兴趣区域ROI  神经网络  MK–Tempotron

英文关键词  dynamic vision sensors  DVS data classification  objects recognition  temporal region of interest  neural networks  MK–Tempotron

中文摘要

      动态视觉传感器(DVS)因其在获取视觉信息时具有低功耗, 低延迟等特性, 本质上十分适用于便携式设备 上的实时动作识别. 在对DVS事件流时域感兴趣区域(ROI)进行定位与分割时, 现有方法往往不能根据不同物体运 动自适应地设定最佳检测阈值、无法对静态场景中少量背景噪声进行过滤, 为此, 提出基于LIF神经元模型和脉冲 最大值监测单元的运动符号检测(MSD), 以实现在多种不同物体运动下事件流时域ROI关键时间点的自适应精确 定位; 在对分类器进行训练时, 对不同的脉冲输入模式, 使用不同的核函数调整突触后神经元膜电位, 使训练得到的 突触权重朝着正确发放的方向改变, 提出一种具有抗噪性的脉冲神经网络学习算法MK–Tempotron. 实验结果表明, 与同类方法相比, 本文方法在DVS数据集上的识别精度能获得高达14.61%的提升.

英文摘要

      Dynamic vision sensors (DVS), due to their low power consumption and low latency when acquiring visual information, are essentially suitable for real-time motion recognition on portable devices. When locating and segmenting the temporal region of interest (ROI) of the DVS event stream, the existing methods often cannot adaptively set the optimal detection threshold according to the motion of different objects, and they cannot eliminate a small amount of background noise in a static scene either. To solve these problems, a motion symbol detection (MSD) method based on the leaky integrate-and-fire (LIF) neuron model and a peak spiking monitoring unit is proposed, to precisely locate the critical time point for the temporal ROI in the event stream containing a variety of different moving objects. And an anti-noise learning algorithm based on the tempotron rule, which is named as MK–Tempotron, is proposed to deal with background noise in static scenes. In MK–Tempotron, different kernels are applied according to different input spiking patterns to adjust the post-synaptic membrane potential of the neuron during training, so that the synaptic weights can be changed in the direction of correct firing of spikes, thus the anti-noise performance of the classification algorithm is improved. Experimental results show that compared with some similar methods, the recognition accuracy of the proposed method on several DVS datasets can be improved by as much as 14.61%.