Fast and accurate object recognition based on tactile data by using Dendrite Net

Tactile perception plays an important role in a variety of applications such as object detection and robotic grasping of objects. However, there is a lack of an intelligent object recognition algorithm based on tactile data with small storage space, fast computing speed, and high accuracy. Inspired...

Celý popis

Uloženo v:
Podrobná bibliografie
Vydáno v:IEEE International Instrumentation and Measurement Technology Conference (Online) s. 01 - 06
Hlavní autoři: Liang, Haonan, Gao, Tianshi, Jiangtao Luo, Tian Gao, Wang, Jiang, Deng, Bin
Médium: Konferenční příspěvek
Jazyk:angličtina
Vydáno: IEEE 22.05.2023
Témata:
Data models
Dendrite Net
Dendrites (neurons)
Feature extraction
object recognition
Predictive models
Real-time systems
Robot sensing systems
robotic
tactile pressure
Training
ISSN:2642-2077
On-line přístup:Získat plný text
Tagy: Přidat tag
Žádné tagy, Buďte první, kdo vytvoří štítek k tomuto záznamu!
Popis
Shrnutí:Tactile perception plays an important role in a variety of applications such as object detection and robotic grasping of objects. However, there is a lack of an intelligent object recognition algorithm based on tactile data with small storage space, fast computing speed, and high accuracy. Inspired by the involvement of dendrites of biological neurons in the computation of neural networks, we introduce dendritic computation to the task of object recognition based on tactile pressure data. Firstly, we design a tactile acquisition glove with a flexible film sensor array based on force-sensitive-resistors to obtain the pressure of the object. We conduct experiments by using this haptic acquisition device and obtain 24,500 tactile pressure images. Then we determine the optimal parameters of the DD model on the tactile pressure data by adjusting the number of DD modules. By using the DD model, we get information about the logical relationships between the tactile pressure image features. Finally, we achieve object recognition by capturing the pressure signal of the object and converting it into a pressure image. We train the model to achieve 99.84% accuracy, while the model achieves a storage space of 1.72MB and a recognition speed of 0.75ms, which outperforms existing algorithms in terms of real-time predictability, model storage space, and accuracy. This paper has significant implications for the future development of haptic applications in robotics.
ISSN:2642-2077
DOI:10.1109/I2MTC53148.2023.10176029