|
| |
|
|
The influence of first CNN layer initialization on training convergence
Krejsa, Jiří ; Věchet, Stanislav ; Chen, K.S.
During evaluation of convolution neural networks on the task of sign language single hand alphabet classification we have discovered that in small but not negligible number of cases the training of the network does not converge at all. This paper investigates the problem that we believe is independent of the application. While the true cause of training divergence was not discovered, we can offer the reader an easy solution from practical point of view – initialization of the first CNN layer using pretrained networks parameters.
|
|
|
Autonomous vehicles lane detection using particle filters
Věchet, Stanislav ; Krejsa, Jiří ; Chen, K.S.
Lane detection belongs among many others crucial tasks which an autonomous vehicles traveling in urban roads need to take care of. We present a lane detection method which uses particle filters combined with visual information from onboard camera to control the vehicles direction. Our initial experiments shows promising results as applied and tested in urban roads with focus on real-time data processing.
|
|
|
Mobile Robot in the Elevator: What Floor Am i On?
Krejsa, Jiří ; Věchet, Stanislav ; Chen, K.S. ; Havelka, M. ; Černil, M.
Wheeled mobile robots in multiple stories buildings have to use elevator to access arbitrary floor of the building. To do so, the control system must be able to detect and access elevator controls and also determine on which floor the elevator stopped. The paper deals with the latter problem, using the fusion of relative floor change detected by onboard accelerometer and absolute floor number detected by the processing of elevator information screen panel image acquired by onboard camera. Bayesian filter is used for data fusion and convolution neural network for image processing. Field tests resulted in 97% correct detection.
|
|
|
Control system architecture for novel experimental microalgae photobioreactor Simbios
Věchet, Stanislav ; Krejsa, Jiří ; Chen, K. S.
Photo-bioreactors for growing biomass based on various microalgae cultures gains interests nowadays due to have possible positive impact to society because of the potential to serve as an atmospheric carbon dioxide removal (CDR) device, produce food or fuel. We present an overall architecture of a control system designed with focus to keep the photo-bioreactor alive for long term continuous operation without any human supervision needed. To test overall control system stability we design a custom build spiral photo-bioreactor (PBR) with novel hyperbolic shaped illuminated part. The shape of the illuminated part was design with the goal to increase the ratio of spiral part to volume due to the fact that the light is crucial for microalgae to grow. The control system was created from scratch while adopting reliable IoT sensors combined with highly reliable custom build embedded control system including a cloud data processing and storage. As a result we have scalable and reliable solution which is suitable for photo-bioreactors of various shapes and sizes to operate in fully autonomous mode without any human supervision needed.
|
|
| |
|
|
Anomalies detection in time-series data for the internal diagnostics of autonomous mobile robot
Věchet, Stanislav ; Krejsa, Jiří ; Chen, K.S.
Autonomous mobile robots are complex mechatronic machines which consists of numerous hardware and software modules working asynchronously to achieve desired behaviour. As there are many frameworks which helps to overcome the flat learning curve the problem of internal diagnostics arises. While users and developers are able to focus only on solving the high level problem algorithm or methods the internal states of the system is hidden. This helps to separate the users from solving hardware issues, which is helping until everything works properly. We present an algorithm which is able to detect anomalies in time based behaviour of the robot to improve the users confidence that the internal robot framework works correctly and as desired. The algorithm is based on probabilistic patterns detection based on Bayesian probabilistic theory.
|
|
|
EXPERT SYSTEM SHELL ARCHITECTURE BASED ON DECISION NETWORK
Věchet, Stanislav ; Krejsa, Jiří ; Chen, K.S.
Presented paper deals with preliminary design of an empty expert system suitable for system monitoring in various engineering applications. Discussed expert system is rule-based with the possibility of importance factor definition within each rule. The deduced result can be composited from more than one possible hypothesis based on different confidence level. The core of the system is based on Bayesian decision network and simple autonomous mobile robot use-case is used for results presentation.
|
|
|
Multiple Indoor Robot Localization Using Infrared Beacons
Krejsa, Jiří ; Věchet, Stanislav ; Chen, K.-S.
Mobile robot localization based on active artificial landmarks is well established robust technique in indoor mobile robotics. This paper explores the extension of the method to multiple robot localization, with the key idea of using the robots themselves as the artificial landmarks, supplying other robots in the group with its estimate of position. The method uses nonlinear version of Kalman filter for determination of position estimate in 2D space, x-y coordinates and angle with respect to global coordinate system, with the estimate represented by mean values and corresponding covariances. Simulation results based on model well verified with the real system suggest that while only small reduction of number of fixed active landmarks can be achieved, the main advantage of the method is in increased robustness of localization technique with respect to obscured landmarks fields of view.
|
host ::
RSS.