| Mobile intelligent robots can help humans complete tedious and even dangerous work.With the more and more common use of indoor service robots,the research on autonomous navigation of mobile intelligent robots has become one of the mainstream directions of robot research.In recent years,the vigorous development of sensors provides innovative ideas for the research of path planning algorithms.Binocular vision is a way to simulate biological perception of the environment to obtain three-dimensional information of objects.The difficulty is the real-time environment perception and path planning algorithm to ensure that the robot can reach the target quickly and smoothly.Path planning of Automated Guided Vehicle(AGV)can be divided into global path planning and local path planning according to their ability to master the surrounding environment information.In this thesis,a hybrid path planning combining global path planning and local path planning is designed.In the initial environment,a global path planning algorithm based on snake optimization is first used to design a global route,and AGV travel along the global route from the starting point.When it encounters dynamic obstacles in the path,it uses Q-learning local path planning for local obstacle avoidance.After completing obstacle avoidance,AGV returns to the previous global path until it reaches the target.The main research contents and achievements of this thesis are as follows:(1)Aiming at the low accuracy of binocular vision stereo matching,an improved census transform stereo matching algorithm based on Otsu image segmentation was proposed.In the cost calculation stage,gray value of neighborhood pixel was used to replace gray value of center pixel of census transform window,which solves the problem of poor noise resistance in traditional census algorithm.The matching accuracy of the depth discontinuous region of the image is further improved.In the cost polymerization stage,the multi-scale cost polymerization method is adopted,which greatly improves the robustness and matching accuracy.In the stage of parallax optimization,the left and right consistency test is first used to detect the mismatched points,and then the region obtained from the previous image segmentation is referred to and filled with the median value of parallax values of nearby pixel points in the region where the mismatched points are located to obtain the optimized parallax map.Through the experimental verification of four groups of standard images in Middlebury website,compared with the traditional census transform algorithm,the improved algorithm improves the robustness and stereo matching accuracy of images.(2)An improved obstacle detection algorithm based on binocular vision k-means image segmentation is designed.for the problem that obstacle detection in binocular vision is easy to fall into local optimal,Firstly,the obstacle determination algorithm based on depth detection is used to evaluate whether the object is an obstacle.Then,the K-means image segmentation algorithm based on bacterial foraging algorithm is used to extract the physical information of the obstacle and eliminate redundant information on the ground of the obstacle to obtain the object area of the obstacle.Finally,the improved census transform stereo matching algorithm was used to obtain the parallax value of the target obstacle.Combining the calibrated internal and external parameters of the binocular camera to perform 3D reconstruction of the target obstacle to calculate its width and height information.The final experimental results show that this algorithm can correctly extract the target obstacle area,and the accuracy of obstacle measurement is improved by 3.5%.(3)Aiming at the problem of path planning in complex environment,this thesis proposes a hybrid path planning that is offline first and online later.Firstly,snake optimization algorithm is applied to the global path planning of AGV to avoid static obstacles and obtain the global path as the initial path.When encountering dynamic obstacles,local online path planning is adopted.In this thesis,a Q-learning AGV indoor path planning method combined with the received signal strength is designed as a local path planning algorithm.In this method,the target transmits signals,and takes the received signal strength as the return function value of Q-learning.To solve the problem that the jitter of path length remains obvious with the increase of iteration times,a state transition strategy was proposed that exploration factor and learning factor change dynamically with algorithm iteration.After successfully avoiding the dynamic obstacle,the AGV returns to the global path until it reaches the target.The final simulation results show that AGV can successfully avoid static obstacles and safely avoid dynamic obstacles.The hybrid path planning based on snake optimization algorithm and improved Q learning algorithm has certain practicability and can be applied to path planning in some complex environments. |
