一种基于触觉信息的脉冲图注意力残差卷积物体检测算法

触觉传感器(柔性电子皮肤)在机器人进行人机交互和工具操作时发挥着重要作用,如何有效利用触觉信息进行物体检测是当前研究的主要瓶颈.鉴于此,提出一种脉冲图残差卷积神经网络SNN-Atten-ResGCN的物体检测算法.首先使用图残差网络ResGCN模型训练触觉时间序列的表征信息,通过引入深度学习模型中的注意力机制拟合触觉数据图形结构的局部特征;然后对重构的触觉图形输入由3个LIF神经元和2个FC全连接层组成的SNN脉冲神经网络训练得到特征向量;最后投票层Vote解码网络特征并检测物体类别.在EvTouch-Objects和EvTouch-Containers两个家庭常见物体触觉数据集上进行对比实验,实验结果表明,所提出方法在保证模型迭代效率的同时,对各种不同的家庭工具对象和容器对象的检测准确率、精度、召回率和$F_1$-score均有提升.

An object detection algorithm based on convolution of attention residuals of pulse graph based on tactile information

Tactile sensors (flexible electronic skin) play an important role in robot human-computer interaction and tool operation. How to effectively use tactile information for object detection is the main bottleneck of current research. Therefore, a pulse graph convolution neural network, SNN-Atten-ResGCN, is proposed for object detection. Firstly, the graph residual network ResGCN model is used to train the representation of tactile time series. Secondly, the attention mechanism in the deep learning model is introduced to fit the local features of the graphic structure of tactile data. Thirdly, the reconstructed tactile graphics are input, and the SNN pulse neural network composed of three LIF neurons and two FC full connection layers is trained to obtain the feature vector. Finally, vote is utilized to decode the network feature components and discriminate object category. Comparative experiments are carried out on the tactile datasets of EvTouch-Objects and EvTouch-Containers. The experimental results show that the proposed method ensures the model iteration efficiency, and improves the accuracy, precision, recall rate and $F_1$-score of various household objects and container are improved.