Online teleoperation of writing manipulator through graphics processing unit based accelerated stereo vision
Remotely operated robots are of great use in unstructured environments where their autonomous counterparts are challenged. These robots are commonly operated using knobs, joysticks or wearable motion sensors. These modes of operation either require a skilled operator or hinder the natural movemen...
Saved in:
| Main Author: | |
|---|---|
| Format: | Thesis |
| Language: | English |
| Published: |
2021
|
| Subjects: | |
| Online Access: | http://psasir.upm.edu.my/id/eprint/103993/1/FADI%20IMAD%20-%20IR.pdf |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| Summary: | Remotely operated robots are of great use in unstructured environments where
their autonomous counterparts are challenged. These robots are commonly
operated using knobs, joysticks or wearable motion sensors. These modes of
operation either require a skilled operator or hinder the natural movement.
Therefore, contactless motion trackers such as stereo vision, structured light and
time of flight systems were invented to allow natural control with minimum
training. Stereo vision systems have the advantage over other systems of being
cheaper while having scalable accuracy and range. These benefits however are
challenged by the high computational requirements of the algorithms used. In
this thesis, a framework is developed to enable the use of stereo vision in realtime
teleoperation of a manipulator robot for the task of writing. The proposed
algorithms aim to accelerate the processes of distortion removal and rectification
using precomputed combined maps. Furthermore, the algorithms avoid
computationally intensive correspondence problem by matching only the points
of interest of a labeled pen to compute its position and orientation. These
algorithms are then parallelized using Compute Unified Device
Architecture CUDA C language to run on Graphics Processing Unit GPU for
hardware acceleration.
Like most sensors stereo camera readings are susceptible to noise due to
lighting conditions and detection errors. Most techniques used in filtering noise
such as the simplest moving average or polynomial spline fitting either eliminates
some original important data along with the noise or are not optimized for realtime
filtering. Thus, Kalman filter which is popular in tracking applications is
proposed. Kalman filter is applied to readings along with the Savitzky-Golay and
Moving Average filters. The performance of filtering methods is compared in term
of processing speed and Root Mean Square Error (RMSE) with ground truth data
collected from a high accuracy digitizing tablet. The proposed technique has
demonstrated a significant improvement by reducing processing time and having
low RMSE that is nearly equivalent to Savitzky-Golay filter and lower than
Moving Average filter. Furthermore, it does not require a window that contains
future data samples.
Cameras capture images at constant frame rates coupled with hand movement
speed variations, this can result in the trajectory points generated have varying
movement distance and exceeding the robot speed constraints. Furthermore,
most manipulator robots in use are designed with the application in mind being
pick and place. Therefore, they follow a speed profile between the pick and place
points that begins with acceleration, cruise and finally deceleration. When
sending a stream of points to robot controller the speed profile results in the robot
having a jittery behavior. To deal with these issues, adaptive sampling algorithm
is applied to the data points to adhere to the manipulators speed constraints.
Then the response of the internal low-level PID controller was examined and
gains were tuned to reduce the rise time to meet real-time requirements. To find
the joint angles from the position and orientation of end effector the inverse
kinematics equations of the arm are computed beforehand. Then joint angles are
streamed at 0.02 seconds interval to ensure that the end effector glides smoothly
along the drawn trajectory.
The performance of the whole system is evaluated by taking the end effector
pose readings from internal sensors through Real-Time Data Exchange (RTDE)
port and comparing those against the data from the digitizing tablet. The results
demonstrated that the movement path is traced by the manipulator while
maintaining minimum error of 0.95 to 1.81mm and minimum delay compared to
other implementations. The positional error is reduced by 52% compared to the
best-known implementation with captured frames being processed in real-time. |
|---|