Research On Technology Of Environment Perception Based On Vision Navigation For Intelligent Vehicles

With the development of society and the progress of technology,Intelligent vehicles have become a popular topic of current research due to the tremendous significance in various fields.The navigation system of intelligent vehicles is made up of environment perception,autonomous location,path planning and decision control from the view of functional structures.Among them,the technology of environment perception,as the eye of intelligent vehicles,is one of the biggest obstacles to commercialization and mass production for intelligent vehicles.In addition,environment perception is the foundation of the implementation of other modules.So far,high-grade driverless vehicle system still depends on the measurement of sensors like laser radar.Unfortunately,the laser radar is not ideal enough in terms of cost and resolution.Therefore,the technology of vision-based environment perception has a rapid development in recent years.Environment perception can be also called environment understanding and scene understanding,it mainly consists of the detection of road drivable region,the detection of lane lines,the semantic segmentation of road scene and the detection of obstacles such as vehicles.In this paper,I devoted myself to analysis the several key technologies and the research status at home and abroad of the vision-based environment perception of intelligent vehicles.I have summarized the shortcomings in the current research of the technology of environment perception.On this basis,I was committed to researching the key techniques of environment perception including the road detection algorithm,the lane line detection algorithm,the image semantic segmentation algorithm and the simultaneously segmentation and detection algorithm.I have proposed several kinds of algorithms to improve the performance of environment perception,which is of great significance and reference value to the development and application of intelligent vehicles and the vision-based navigation system.Generally,this paper mainly finished works as follows:(1)Research on road detection algorithm of environment perception for intelligent vehicles.A new algorithm which fusing road appearance cue and prior cue was proposed in this paper.Firstly,input images are preprocessed by SLIC,morphological processing and illuminant invariant transformation to get superpixels and remove lane markings,shadows or highlights.Then,we design a novel seed superpixels selection method and model appearance cue using Gaussian Mixture Model with the selected seed superpixels.Next,we propose to construct road geometric prior model offline,which can provide statistical description and relevant information to infer the location of the road surface.Finally,Bayesian framework is used to fuse appearance and prior cues.The proposed algorithm of fusing appearance and prior cues can provide reliable road detection results and show robustness to various driving scenarios,shadow,highlight,lane lines,obstacles and so on.Experiments are carried out on KITTI road benchmark where the proposed algorithm show compelling performance with MaxF value of 92.51%,which achieves state-of-the-art among the training-free methods.(2)To tackle the problem that current lane line detection algorithms usually have insufficient robustness and can only detect a fixed number of lane lines,which resulted in the intelligent vehicles can’t deal with the problem of changing lanes.I carried out research lane line detection algorithm of environment perception for intelligent vehicles.A new multi lane lines detection algorithm based on FCN and H-Net was proposed in this paper.Our method regards the question of lane line detection as the question of image semantic segmentation.This algorithm segments the input image into two semantic classifications including background region and lane line region by designing a lane line segmentation network based on FCN.Then,it generates a perspective transformation matrix and perform perspective transformation based on the segmented image.Next,perform curve fitting to the lane line pixels obtained by the above steps.Finally,obtain the mathematical model by carrying out inverse transformation to the original image.Additionally,define the region between the nearest two lane lines as ego lane of intelligent vehicle.Experiments were carried out on TuSimple dataset where the proposed algorithm has a accurate rate of 96.89%,which outperforms other lane line detection algorithms.(3)Research about image semantic segmentation algorithm of environment perception for intelligent vehicles.A kind of semantic segmentation algorithm based on improved DeepLabV3+ and refinement by superpixels was proposed.To tackle the problem that DeepLabV3+ semantic segmentation algorithm can’t recover sufficient details of input image by using bilinear upsampling with the factor of 4 twice,this method optimized DeepLabV3+ by using deconvolution.Then,semantic features were extracted and coarse semantic segmentation result was obtained by the refined DeepLabV3+ network.Additionally,segment the input image into superpixels by SLIC superpixel segment algorithm and fuse the abstract semantic features and the details information of superpixels.Finally,semantic segmentation results of edge refined was obtained.This algorithm can deal with the problem that a large amount of detail information was lost by using pooling layers and down-sampling layer,which lead to the terrible performance at the part of image details like edges.Experiments on PASCAL VOC 2012 dataset and Cityscape dataset show that my proposed semantic segmentation algorithm shows better semantic segmentation performance in the part of image details like edges than DeepLabV3+.The mIoU of my proposed algorithm reaches 82.8%,which achieves state-of-the-art among the image semantic segmentation algorithms.(4)Research on the technology of multi-task learning and the problems need to be tackled in the field of environment perception for intelligent vehicles.A new algorithm that simultaneously perform segmentation and detection was proposed.This algorithm has a Encoder-Decoder architecture whose encoder is shared by the segmentation network and the detection network.The segmentation part of the decoder obtains the image road region segmentation result.Correspondingly,the detection part of the decoder obtains the image object detection result.Experiments carried out on KITTI dataset show that the algorithm proposed this paper can simultaneously obtain the segmentation to drivable region of road the detection to vehicles in the image by using one network.The proposed method tackled the problem that the technology of image semantic segmentation can’t distinguish different instances of the same semantic classification and the technology of image object detection can’t detect the road region by using bounding boxes.The algorithm proposed in this paper can improve the efficiency of environment perception algorithms,which can accomplish the complex task of environment perception by using one network.