Research On The Key Technologies Of Visual Perception On Driving Environments Of Intelligent Vehicles

Driving environment perception is the key technology of unmanned vehicle,and it is the premise and foundation of decision-making and control system.Accurate,reliable,real-time environmental perception is the guarantee of intelligent vehicle driving safety.Based on the machine vision,the traffic environment perception technology has the advantages of low hardware cost and more image information.Especially,the deep learning has made breakthroughs in image processing and recognition in recent years,and provided a new environment perception solution for the vision system of intelligent vehicle.Therefore,in view of the common problems of image feature extraction and classifier design in vision perception system,this thesis studies the key technology of visual perception in a variety of traffic participation elements in intelligent vehicle driving environment,including:（1）A pavement detection algorithm based on full convolution deep neural network model and conditional random field is proposed.Firstly,the VGG₁6 deep convolution neural network with 13 convolution layers and five pooling layers is used as the basic network,and a set of three different spans of the up-sampling prediction output layer is designed to achieve the pixel level road segmentation image output with the same size as the original input image.secondly,the algorithm combines conditional random field with the convolution neural network,then the road edge from the up-sampling prediction output layer is optimized.The test results show that the proposed method is a more effective road detection method,which is about 6% higher than that of the Nearest Neighbor method,about 4.5% higher than the K-means method and about 4.1% higher than the Robust Gaussian method in accuracy.（2）A structured road marking detection and recognition algorithm based on pavement priori knowledge is proposed.Firstly,based on the results of the priori road pavement,the color contrast is used to define the saliency of each pixel in the image.Through the sliding window search of different dimensions and scales,the region of interest is determined finally.secondly,the VGG₁6 deep convolution neural network architecture is used to achieve six kinds of road marking identification,including the deceleration line,the guiding arrow,the parking line,the no-turn mark,the roadway width reminder line,and the speed limit line.The results show that the proposed method of road marking detection combined with the significance and deep learning is about 20% higher than that of the traditional method,and the average running time is reduced from 90 ms to 52 ms,which can meet the requirements of complex road marking and identification of intelligent vehicle under complex driving environment.（3）A vehicle target detection algorithm based on hierarchical saliency cyclic convolution neural networks is proposed.Firstly,based on the VGG trained model,the saliency of input image is detected and the coarse salient map of input image is accomplished.Then based on the hierarchical recurrent neural network,the last layer of the salient map is refined layer by layer to obtain a more detailed salient map.Secondly,the adaptive threshold segmentation method is used to get the results of the salient target region.The results show that the proposed algorithm obtains accuracy rate of 0.91,recall rate of 0.88 and F-value of 0.89 on the KITTI database,which can provide an effective way for the intelligent vehicle to perceiving the vehicle target.（4）A traffic sign recognition method based on HOG-Gabor and Softmax is designed.Firstly,to extract the HOG feature and the Gabor feature,and then the extracted HOG and Gabor eigenvectors are directly connected in series,and the amplification vector of traffic signs is obtained.Secondly,the softmax classifier is used to classify the amplification vectors to realize traffic sign recognition.The results show that the proposed algorithm obtains the correct recognition rate of 97.68% and the real-time performance of 0.08 ms for each image.Therefore,the correctness and real-time of the method are improved remarkably.（5）A pedestrian detection method based on D-S evidence combination theory is designedFirstly,the HOG feature and the LBP feature are extracted from the input image,and then the pedestrian posterior probability is obtained by probabilistic SVM based on the above two features and the independent evidence is obtained at the same time.Secondly,the basic evidence of each independent evidence is obtained according to the posterior probability.Using the DS evidence combination algorithm based on matrix analysis,all the basic reliability of fusion and output as a new trust function value.Finally,based on the trust function,the algorithm develops decision rules to achieve pedestrian detection or not.The experimental results show that this method can effectively realize the fusion of pedestrian HOG and LBP feature at the decision level,and can make full use of the HOG feature and LBP feature,which is good at characterizing the pedestrian contours and textures of pedestrian appearance separately.Therefore,the proposed algorithm has better pedestrian detection performance.The above methods were tested on the state of art data set,and the results show that the proposed methods can provide an effective solution for the environment perception of intelligent vehicle driving.In this thesis,a variety of exploration to the environment participation detection based on deep convolutional network,this thesis tries to excavate and reveal the learning mechanism of the deep convolution network for characteristics of internal target,and provide a variety of effective and reliable algorithms for the automatic driving task based on the deep system,which has certain potential practical value.