Research For The Road Scene Recognition And Path Planning For The Outdoor Mobile Robots

As the widespread use of the mobile robot,its varieties and application environments are increasing and expanding from indoor environment to outdoor environment.For autonomous mobile robots in outdoor work environment,including driverless cars,the better abilities of environment apperception and scene understanding are the prerequisites for the autonomous navigation and completing its action task.Scene recognition for complex outdoor road environment has become one of the hot research field of the mobile robot.For driverless cars,the biggest challenge comes from the environment apperception,which is the most basic part of the self-driving technology.Due to the complex and varied outdoor road environment,and the instability of the environment information obtained by the autonomous mobile robot,it is difficult for the robot to identify the road and objects,especially under some environment such as varying illumination,pedestrian,traffic jam,suburban and park road without obvious lanes and markings,or rural road without better road surface and obvious edge.In complex environment,how to robustly identify and understand the content of road scenes in real time is a challenging research topic at present.The outdoor road structure can be divided into structured road(highway,urban road),quasi-structured(suburb road,park)and unstructured road(country road,off-road environment).Whether the driverless car can accurately perceive the surrounding environment information of the vehicle in the various road environment is related to the safety of self-driving,pedestrians and other vehicles.Because of the diversity and complexity of road environment,the development of percept technology has not reached sufficient safety at present,so there are still a lot of works to do.First,aiming at the environment perception problems in all kinds of road environment,this paper does some research in road vanishing point detection and road detection methods,and a fast line space voting method for the road vanishing point and road detection is presented,which provides the correct driving direction and the drivable road area for driverless cars in real time with fast speed,high accuracy and good adaptability.Secondly,for the structural environment of city roads and highway,we research the road scene understanding method based on the deep learning,and a road scene segmentation network based on the deep convolutional neural network is presented to identify the road area,the pedestrians and vehicles on the road,sidewalks and other category of objects effectively,and provide diverse road scene information for driverless cars,which enhance its ability to identify environment.Then,aiming at the problem of multi-sensor information processing,we research the multi-information fusion methods for road scene understanding and present a multi-information fusion framework based on belief function theory in order to fuse detection results of the multiple sensors to improve the accuracy and reliability for road scene recognition.Finally,according to the drivable road area and the target point detected by the perceptual system,we study the mobile robot local path planning method in dynamic environment,and an improved teaching-learning-based optimization algorithm is presented and applied to the path planning problem of the mobile robot,which provides a new theoretical method and practical basis for the mobile robot path planning.The main contents of this paper are as follows:1.The research for fast detection of the road vanishing point and road detection based on the vanishing point.Analysis of recognition method in all kinds of unstructured road and quasi-structured road environment,according to the texture and color characteristics of the road image,a linear space voting method based on texture direction to estimate the road vanishing point is presented.Based on the detected vanishing point,the two main boundary lines of the road are detected to get the road area by using the voting ratio of interval bin and the color difference of interval bin.2.The research for road scene understanding based on deep learning.For the structured environment of the urban road and highway road,a deep road scene segmentation network composed of 16 layers convolutional encoded-network and 16 layers deconvolution decoded-network based on the convolution neural network is presented.By increasing the number of network layer,changing the nonlinear mapping function of the activation layer and upsampling method,and using fine-tuning strategy of transfer learning to build a road image segmentation model which can improve the training speed and classification accuracy,thus the road scene image can be semantically segmented effectively,even scene content can be precisely classified for recognition at the pixel level.3.The research of multi-information fusion for road scene understanding.By analyzing the correlation between the information gathered by the sensors,we study the multi-information fusion method for road scene understanding,combine and model the output results of a variety of detection modules based on the theory of belief function,and build a multi-information fusion framework for road scene understanding.Multiple detection modules are used to process the data of sensors,and the detection results of each module are fused in the same space to improve the accuracy and reliability of the road scene understanding.4.The research and improvement of the teaching-learning-based optimization algorithm.By studying the teaching-learning-based optimization algorithm(i.e.a new swarm intelligent optimization algorithm),a nonlinear inertia weighted teaching-learning-based optimization algorithm is proposed to solve global optimization problems in continuous space.And the proposed optimization algorithm and other classical optimization algorithms are compared on many benchmark functions by a number of experiments.5.The research for local path planning of mobile robots based on nonlinear inertia weighted teaching-learning-based optimization algorithm.We study local path planning method of mobile robots in dynamic environment by combining with the improved teaching-learning-based optimization algorithm.Through modeling the environment map,this improved algorithm is applied to the local path planning of mobile robots to find the optimal and drivable path on the map.