Mining local and global spatiotemporal features for tactile object recognition.

Area of Science:

• Robotics
• Artificial Intelligence
• Computer Vision

Background:

• Tactile object recognition (TOR) is crucial for robot environmental perception.
• Existing methods using single-scale convolution struggle to extract both local and global spatiotemporal features from tactile data, limiting TOR accuracy.

Purpose of the Study:

• To propose a novel network architecture for enhanced tactile object recognition.
• To improve the accuracy of TOR by effectively capturing multi-scale spatiotemporal features.

Main Methods:

• Introduced the Local and Global Residual (LGR-18) network, featuring multiple Local and Global Convolution (LGC) blocks.
• Each LGC block integrates Local Convolution (LC) modules (using temporal shift and 2D convolution) and Global Convolution (GC) modules (fusing 1D and 2D convolutions).
• The LGR-18 network extracts local-global spatiotemporal features without relying on computationally expensive 3D convolutions.

Main Results:

• Ablation studies confirmed the effectiveness of the LC module, GC module, and LGC block.
• Quantitative comparisons demonstrated superior performance against state-of-the-art methods in TOR tasks.

Conclusions:

• The proposed LGR-18 network effectively extracts local-global spatiotemporal features for tactile object recognition.
• This approach offers a parameter-efficient alternative to 3D convolutions, achieving excellent performance in TOR.