Gradient Adaptive Sampling And Multiple Temporal Scale 3D CNNs For Tactile Object Recognition

Tactile object recognition is one of the key tasks in robotic tactile perception,aiming to identify the objects grasped by the robot.With the rapid development of artificial intelligence technology,deep learning-based tactile object recognition methods have achieved better recognition results than traditional methods,but there are still three problems that need to be solved in such methods as follows:(1)the existing methods use a uniform sampling strategy that can easily cause redundancy or missing data;(2)the existing methods do not have enough generalization ability for tactile data with different acquisition frequencies;(3)The existing methods directly use the visual depth model as the backbone network resulting in reduced feature extraction capability.To address the problem(1),this paper proposes a gradient adaptive sampling strategy,which can adaptively select key data from the acquisition data for subsequent object recognition.The gradient adaptive sampling strategy measures the importance of a tactile frame according to its gradient value and adaptively determines the sampling interval in order to obtain as much critical tactile information as possible when the same number of input frames.To address the problem(2),this paper proposes a multiple temporal scale 3D convolutional neural network model,which can extract and fuse depth features at different temporal scales to recognize object categories.The multiple temporal scale 3D convolutional neural network acquires tactile data at multiple temporal scales through multiple temporal scale downsampling,and then recognizes objects by extracting and fusing multiple temporal scale deep features,and the fused features have better generalization ability for different acquisition frequency tactile data.To address the problem(3),this paper improves the existing lightweight 3D convolutional neural network ResNet3D-18 and constructs the MR3D-18 network.The MR3D-18 network removes a pooling layer and adds a dropout layer from the ResNet3D-18 network,which can improve the matching of the network to small size tactile data on the one hand,and prevent the overfitting of the network on the other hand.The ablation experiments on the MR3D-18 network show that the proposed MR3D-18 outperforms the original ResNet3D-18 network.The ablation experiments on the gradient adaptive sampling strategy and the multiple temporal scale 3D convolutional neural network show that both the gradient adaptive sampling strategy and the multiple temporal scale 3D convolutional neural network can effectively improve the object recognition accuracy.The comparison experiments between the methods of this paper and popular methods on two publicly available tactile datasets show that the overall accuracy of the proposed method is optimal,which fully validates the effectiveness of this paper’s method.