| As multimodal camera networks have been deployed in various environment, image fusion is playing a critical role for better visual perception and process parameter measurement. The objective of the dissertation is to design robust vision-based thermal control systems to tolerate uncertainties for industrial automaton. To be specific, two new methods have been proposed, one for robust shape fitting in visual images and another for packet loss recovery in thermal images.;Secondly, the thermal image loss generated by network traffic from camera nodes to fusion center is modeled as a Markov chain. A graph cuts method is proposed to recover the loss based on thermal pattern classification. Simulation results show that thermal information can be partially retrieved, which may greatly increase the robustness of a thermal management system.;The proposed methods are tested with a laboratory die casting process simulator with two visual cameras and one thermal camera. A simple fuzzy PID controller is designed to integrate the visual sensors into a control loop. The experimental results show that the homogeneity of the temperature distribution in a die may become achievable through the vision based thermal control system.;Firstly, an adaptive curve fitting technique is proposed based on prediction error sum of squares for the sampled data set containing outliers. The method converges very fast and can achieve superaccuracy under certain conditions when compared with other algorithms. The method is applied to estimate an optimal curve equation of a casting die in the visual images. |
