Biography
Enrollment Date: 2014
Graduation Date：2017
Degree：M.S.
Defense Date：2017.05.24
Advisors：Chun Zhang
Department：Institute of Microelectronics，Tsinghua University
Title of Dissertation/Thesis：Design of Bimodal Tactile Sensor
Abstract:
With the rapid development of robot technology, robots are transfering from an era of industrial robots to an era of the intelligent robots. The main difference between industrial robots and intelligent robots is that the latter is integrated with many sensors ,which enables it the ability of performing muti-modal information fusion and more elaborate actions. Tactile sense which contains perceptions of force, texture, temperature and so on is an important basis for lots of complex operations. According to the research, most commercial sensors in use are single-modal, which means that they can only measure one physical quality and thus split the relationship between various physical quantities. In conclusion, the demand for muti-modal sensors is increasing.
The tactile sensor we designed is based on visual information which integrates force detection and texture recognition by image processing. First of all, we design a set of hardware structure, which contains mechanical shell, lighting circuit and the elastomer. Of which the elasomer is the most important, its basic component is transparent PDMS, whose softness is similar to the fingers of a man. In order to detect force .markers are made on the surface of the elastomer by sputtering and laser method. At the same time, with the purpose of integrating markers and texture, another layer of copper is sputtering outside of the markers.
This thesis also introduces an algorithm to identify markers and recognize texture type, the key of which is to eliminate the mutual influence between these two modals. The methods of marker identification includes two parts: NCC template matching and centroid of connected field. The first method is more accurate, but it also needs large amount of calculation, which is more suitable to be implemented in computer. While the second method contains binarization, calculating connected domain and centroid, whose calculation is simpler and more suitable for hardware implementation. According to the marker positions of each frame we can get the changing torque vector and finally get the changing force. On the other hand, texture recognition is achieved by LBP operator, which has the ability of rotation and gray-scale invariance. First we need to build the texture database.
In order to increase the flexibility of these sensors, we also introduce the algorithm implementation on the hardware platform in this paper. We design two systems based on the FPGA platform, one is an SOPC system using NIOSII CPU and the other is made of pure logic circuits. The first system controls DMA to read data from FIFO and then store them in SDRAM. All the algorithm is implemented is CPU and the output image is shown on the screen . In the second system, image capturing, data preprocessing and the realization of the algorithm are all implemented using pure circuits, which greatly improves processing speed.
At the end of this thesis, we introduce the 3-d reconstruction. The innovation is that our algorithm is able to deal with shadows. On the basis of traditional photometric stereo method, we also add the shadow processing part, which improves the precision of the reconstruction algorithm greatly.