The Research On Methods Of Mobile Robot Particle Filter Localization And Mapping

With the development of computer science, sensor technology, artificialintelligence and other disciplines, and increasing levels of mechanical design andmanufacturing, the mobile robot is increasingly toward the development of intelligentand autonomy-oriented. Autonomous localization and map building is the basis of thestudy of the mobile robot intelligent navigation and environmental exploration, andlocalization accuracy and the accuracy of the built map is a prerequisite for its abilityto successfully applied in a real environment. The working environment of the mobilerobot can be divided into two kinds of outdoor environment and indoor environment,this paper studies mobile robot self-localization and mapping in the indoorenvironment.As a result of the vision sensor rich information, vision sensor is suitable formobile robot localization based on the map. But the computational complexity ofvisual information, perspectives change, environmental light changes make visualfeature extraction and matching becomes difficult, the ability to find accurate andefficient feature extraction and matching is the key map-based localization methods.The ability to find an accurate and efficient feature extraction and matching methodwhich is the key of the robot localization based on the map. The scale invariantfeature transform (SIFT) for image scaling, rotation, three-dimensional perspective,noise, light intensity changes with better invariance, is a very reliable feature pointdetecting method, widely used in mobile robot localization and mapping area. A largenumber of redundant key points which are not required for feature matching areproduced in the process of feature extraction, and the computational of SIFT featuresis too complex. Therefore, it is necessary to reduce the number of key points at thesame time does not affect feature matching. For robot vision localization problem, toimprove the classic SIFT algorithm to exhaustive search method, a localizationmethod based on the iterative SIFT Monte Carlo localization (ISIFT-MCL) isproposed. The ISIFT-MCL method is invariance to changes in light intensity, imagescaling, three-dimensional perspective and noise, but also to reduce SIFT algorithmproduces unnecessary feature point, to reduce extraction and matching time of thefeature point. Environment feature information and odometer information via particlefiltering integration, obtaining more accurate environmental mark point coordinate. During the mobile robot localization, the perception of the environment feature pointinformation and odometer are fused through the particle filter, to obtain a moreaccurate environmental labeling coordinates.In most cases a single sensor can only provide part of the outside information, themobile robot is difficult to make an accurate judgment based on single sensorinformation. The use of information fusion technology to all kinds of multi-sensorcomprehensive information, to get complementary, redundancy, completeenvironmental information, to make robots robust to localization. In order toeffectively use perception information of visual and laser ranging, taking intoaccount the particle filter of the monocular vision sensor and laser rangefinder, afusion of laser ranging information and visual information of the particle filterlocalization method of mobile robot is proposed. Firstly, this paper proposes amapping method based on adaptive curvature calculation. Split from the laser scandata to the three features: straight line segments, curve segments and corner point, tocreate the map of environmental features by extraction and description of alreadysegmented feature. Secondly, this paper proposes a new method of image retrievalwhich can be used for robot localization. HSV color model and Gaussian mixturevector quantization (GMVQ) are selected to extract the color histogram of the imageof an environment. These images are collected during robot training is stored in thedatabase. Each of images in the database constitutes a specific class, to associate withthe images in the database and environmental locations. Using the Euclidean distancefor image similarity measure, the robot collection of current environmental image andthe images in the database for matching, retrieval the most similar images to thelocation environment for robot vision localization. Finally, a robot particle filterlocalization new method is proposed based on laser scanning and image retrieval. Inthe update phase of perception, this paper uses probabilistic methods for theintegration of perceptual information. The uncertainty of the distance measuringsensor is corrected with the vision sensor information, and the slow vision sensoridentification information is made up with the ranging sensor information. The robotis better adapted to the complexity of the indoor environment.In the localization process, the robot’s uncertainty and unpredictability of thesurrounding environment makes localization difficult. Probability theory can bettersolve the problem of mobile robot localization, but the existing particle filteralgorithm requires a large number of particles to describe the posterior densitydistribution. Importance resampling causes particles set to contain many repetitive particles, causing particles depleted. According to the existed problems of particlefilter algorithm, this paper proposes a robot particle filter localization method basedon interval analysis. Interval analysis can well deal with Gauss white noise andmeasurement error. Multiple particles continuously distributed throughout theinterval, the weight of the particles is not in the traditional sense of the intervalcorresponding to the respective particles, but the corresponding interval weight. Theinterval of particle filters is mainly related to the interval arithmetic, constraintsatisfaction processing, and so on. Interval analysis solves the problem which theparticle filter requires a large number of particles, to make the particles movetowards the real mobile robot trajectory, minimize the required number of particles toachieve the same accuracy requirements.FastSLAM algorithm recursively estimates robot pose and landmarks withRao-Blackwellised particle filter. In the FastSLAM algorithm, state estimation usingparticle filter algorithm, maps are created by N independent EKF (Kalman Filter,EKF) algorithm. The EKF is used to first-order linearization the nonlinear system togenerate a large number of linear errors; at the same time there are also the particlesdepleted and degraded. In view of the above problems, this paper presents asimultaneous localization and mapping method of FastSLAM framework based onparticle swarm optimization. Particle swarm optimization is introduced to the particlefilter algorithm, using the particle swarm optimization particle filter to estimateparticle posterior pose proposal distribution, to solve the problem which the particlefilter needs a large number of particles and particles depleted and degraded. In thispaper, the unscented Kalman (UKF) is used for the landmarks position estimation andmap updated. The experimental results show that the proposed method can get higherprecision of localization and mapping method than the FastSLAM in the case of thesame number of particles.In the end, the whole paper is summarized, and the prospective researchingfields are discussed.