Study On Road Scene Perception Algorithm Based On Convolutional Neural Network

In recent years,with the development of deep learning technologies,amazing advances have been made in such fields as computer vision,natural language processing and automated vehicles.Many enterprises and research institutions are actively studying the technology of automated vehicles.Autonomous vehicles technology can be divided into three parts: environmental perception,planning and decision-making and motion control.Among them,perception technology is the key to achieve this technology.The perception of environment is directly related to the implementation of follow-up function modules.The main task of unmanned vehicles perception system is the perception of road scenes.As one of the key issues in computer vision,scene perception is still not solved completely.The complexity and diversity of road scenes make the perception of road scenes more challenging and difficult.In this thesis,we have carried on thorough investigation and research for road scene perception system.In view of the current high cost of road scene perception system,this thesis decomposes the road scene perception task into two parts with the aim of low cost solutions.One part is the semantic segmentation of road scenes to achieve the simultaneous detection of multiple objects such as pedestrians,vehicles and roads,the other is the unsupervised monocular depth estimation of road scenes.In order to achieve the semantic segmentation of road scenes,we studied a variety of semantic segmentation models based on convoluti onal neural networks,and compared the performance of different models by experiments.In this thesis,we designed a semantic segmentation model based on fully convolutional neural network.Our model applies the unpooling method which does not need learning to upsample.We designed skip-architectures to combine the local features,which makes the segmentation precision been improved.In order to meet the requirement of effectiveness both in terms of memory and computational time during inference for road scene perception applications,we optimized the structure of convolutional neural network to make it lightweight.Based on the residual units and dilation convolution,we constructed a new model with good real-time,and combined different upsample methods to ensure the segmentation accuracy.Collecting a large number of quality depth data in a series of environments is a cost and challenging task,this thesis studies an unsupervised monocular depth estimation method of road scenes.We pretend the depth estimation problem as an image reconstruction problem.This thesis applies the convolutional neural network techniques and combines with disparity consistency constraints and multiple losses to train to generate disparity map without using ground truth.According to the epipolar geometry,the depth information can be obtained given the baseline distance and the camera focal length.In this thesis,we trained and tested the proposed methods on a number of published road scene datasets.In addition,we also collect ed and produced our own datasets to verify the effectiveness of the proposed models.