Route Searching Based On RBF Network And Q-Learning And The Mobile Guidance System For The Visually Impaired Individuals

Due to the limitation of a civilian Global Navigation System(GPS)accuracy,it is difficult to detect the position of an object during complex road environments in real-time.With this lacking capability of route searching and planning for visually impaired individuals,accuracy is more difficult to achieve for safe guidance.Therefore,introducing an intelligent learning algorithm for the navigation system to improve the capabilities of route searching and planning is critical.From further research,the Q-Learning algorithm and RBF network have showed excellent performance in accurate positioning.In addition,some issues occur that can inhibit efficient performance.Poorly designed GPS with our current infrastructure adds on another layer of obstacles that needs to be considered for aiding the visually impaired individuals.Furthermore,the existing GPS equipment for the visually impaired individuals poses disadvantages of single functions and exorbitant prices(such as the visually impaired individuals' sticks and the guiding instrument for the visually impaired individuals).The built-in multi-mode sensors of modern smartphones(such as camera,microphone,infrared sensor and gyroscope,etc.)can be utilized to detect the road information.It is an opportunity to develop a mobile guidance system for the visually impaired individuals with the popularity of smartphones.In this dissertation,the optimal route searching and the design of mobile guidance system were studied and the specific research content is as follows:Firstly,an improved and much stronger RNH-QL method based on RBF network and heuristic Q-learning was put forward for route searching in a larger state space.The new RNH-QL method solves the problem of inefficiency of reinforcement learning.If a given problem's state space is increased,there is a lack of prior information on the environment.The gradient descent method was adopted as RBF network's weight updating rule,thus,reward shaping can give an additional feedback to the agent in some intermediate states.This helped to guide the agent towards the goal state in a more controlled fashion.Meanwhile,In the process of Q-learning,it is accessible to the underlying dynamic knowledge with the background knowledge of an upper level RBF network.It improves the learning efficiency by incorporating the greedy exploitation strategy to train the neural network,which has been testified by the experimental results.Secondly,an identification system of the visually impaired individuals' road and zebra crossing based on the Android platform has been established in this dissertation.The units of image acquisition and image processing are the core of the system,the identification of the visually impaired individuals' road and zebra crossing have been achieved.The identification system can be divided into the pretreatment unit,the morphological processing unit,the coarse positioning unit and the fine positioning unit.The image segmentation,the corner detection and clustering analysis algorithm have been improved for the field of the identification of the visually impaired individuals' road and zebra crossing.The chromatic visually impaired individuals recognition algorithm based on the Otsu threshold segmentation algorithm in the color space of Lab(Lab-Otsu)and the method of Harris corner detection(LO-Harris)are applied to the visually impaired individuals' identification and the detection of visually impaired individuals' corner which can effectively improve the accuracy of visually impaired individuals' positioning.Aiming at the achromatic visually impaired individuals' road,this dissertation introduced the LBP operator into the area of visually impaired individuals' recognition.By combining LBP operator with the K-clustering analysis and subsequent filling operation,the achromatic visually impaired individuals' recognition algorithm(SLBP-KB)has been improved.Furthermore,the Canny edge detection and Hough cumulative transformation have been used to identify the zebra crossing.The method of calculating the moment of the image to detect the contour of the obstacle in the image helped distinguished basic contour edge of the obstacles.Thirdly,a new type of mobile guidance system based on smart phones for the visually impaired individuals have been built.The system includes global navigation system,integrated image processing system and voice auxiliary system.The real-time GPS signal is derived from the basic navigation system and the signal could be processed by the phone's built-in navigation software to obtain the location information.When the destination information has been inputted by the visually impaired individuals' voice,a path would be planning by the system.Comprehensive image processing system includes a visually impaired road recognition module and a traffic intersection recognition module.The real-time information of the visually impaired road could be captured by the camera of smartphone in the visually impaired road recognition module,the function of locating the visually impaired individuals area can be achieved through the processing.The information of traffic crossing could be obtained by the sensors of smartphone in the traffic intersection recognition module,the ability of locating the zebra crossing is beneficial to identify the information of road when closed to the crossing.The processed results of each system could be converted to the corresponding voice prompt,which could guide the visually impaired individuals in the voice auxiliary system.Aiming at the mobile guidance system for the visually impaired individuals which was designed.The optimal route searching,the identification of the visually impaired individuals' road,the identification of zebra crossing and the identification of obstacles have been studied respectively.These studies provide a directional guide for the deep research and the development of the guidance system for the visually impaired individuals,which has the certain theoretical and applied value.