Inteligentno upravljanje mobilnim robotima na osnovu neuro-fazi-genetskog prepoznavanja objekata i praćenja ljudi u robotskoj viziji

Abstract

Robot Vision system is lately an important part of the robot system and it becomes main source of the data used for robot environment perception in order to adequate interaction between robot and surrounding objects and humans. There are numerous advantages of the vision system, but there are also disadvantages like unrobustness of the vision system considering external influences, very complex analysis of cluttered scenes or object recognition when object shape changes. Since current state of the art shows growing potential of development, improvement and implementation of robot vision system based on artificial intelligence and machine learning, for object recognition and human tracking, this field of scientific research is used as a basis for research presented in this doctoral dissertation. During the research it was mandatory to develop vision system that collects data and enables robot to reliably and autonomously function in dynamical environment. Vision system needs to enable feature extraction and classification of different objects from the scene without large databases commonly used in modern robot systems. Therefore, several new algorithms based on neural networks, fuzzy logic and genetic algorithms (as well as some other artificial intelligence algorithms) were developed and used in intelligent mobile robot platform for object and human detection, tracking and interaction. Presented research is guided towards development of mobile robot platform and its control system based on intelligent robot vision algorithms for object and human recognition (classification) and tracking in lab environment. Initial research were done with robot platform where vision system is used as main sensor for manipulator control, and developed algorithms were basis for development of algorithms for human detection and tracking by mobile robot. Initial research was done in order to enable object recognition for DaNI mobile robot platform with monocular camera and FRIEND robot system with manipulator equipped with stereo camera. For adequate interaction of robot with surrounding objects segmentation algorithms and intelligent classification algorithms were developed. Further research was done for human tracking, so experimental mobile robot platforms with stereo camera, thermal camera and RGBD sensor were developed. In this dissertation improvement of robot vision module by neural network classificatory is presented as well as concept of GA-optimal neural network input selection. Implementation of thermal camera for human tracking into mobile robot system is not very common, so development of new fuzzysegmentation algorithm and new genetic algorithm for optimal segmentation parameter tuning of thermal image was done. Also, software based on Support Vector Machne (SVM) was developed for classification of objects in thermal image (human detection in thermal image). As an improvement of human tracking algorithm, in this dissertation nonlinear autoregressive neural network is suggested for prediction and estimation of human position filmed by stereo camera. Finally, some solutions for intelligent control of mobile robot platform equipped with RGBD sensor for human tracking were suggested and with adequate adaptation this platform can be used for gesture recognition. Results presented in this doctoral dissertation confirm that implementation of modern methods and algorithms from the palate of artificial intelligence and machine learning tools allows development of vision based control system that enables mobile robot to detect, recognize and locate objects and humans, that further enables intelligent interaction of robot system with surrounding environment. Implementation of more than one artificial intelligence technique (like artificial neural networks, fuzzy logic and genetic algorithms) into intelligent control system development enables autonomous and reliable object and human recognition and tracking that is superior compared to current solutions.