Scene Understanding And Perception In 3D Environment

Unmanned Autonomous Systems(UAS)is an important support system for unmanned autonomous devices such as unmanned ground vehicles and unmanned aerial vehicles to operate smoothly in real scenarios.It has been widely used in various production and living scenes in society,such as transportation,emergency rescue,aerial photography,etc.Therefore,it is of great theoretical significance and application value to realize the perception and understanding of unmanned ground vehicles visual scene and unmanned aerial vehicles visual scene of the unmanned autonomous system.In order to achieve the unmanned autonomous system of unmanned ground vehicles visual scene and unmanned aerial vehicles visual scene perception and understanding of the visual scene,this thesis studies the unmanned autonomous intelligent perception module in the system,and designed a to the scene to conduct a comprehensive understanding of technical scheme,the scene understanding task is divided into semantic segmentation,target detection and depth estimation and target distance measurement are three related subtasks,and then realize the semantic information of the scene,the location of the object information and the depth of the target information,such as the distance information access,which provide reliable scenario for unmanned autonomous system information.Although domestic and foreign scholars have made a series of achievements on these three sub-tasks,there are still the following problems: 1)The existing semantic segmentation algorithms based on the visual scene of UGV often fail to give consideration to both real-time and accuracy;2)The semantic segmentation task based on UAV visual scene has few baseline data sets,especially the high-resolution UAV visual image data sets;3)In the aspect of target detection,the existing algorithms have poor detection effect in the field of vision restricted area and boundary area;4)At present,the accuracy of the depth estimation model in the edge and detail is low,which brings great challenges to the distance measurement.Through in-depth analysis and summary of the deficiencies and challenges of the existing algorithms,this thesis carried out research work in the following four aspects:(1)Semantic segmentation of visual scenes of unmanned vehicles is achieved.This thesis proposes a lightweight semantic segmentation model for the visual scenes based on unmanned ground vehicles,that is,a asymmetric bottleneck network model.In order to achieve faster model inference speed,the mainstream encoder-decoder structure is abandoned in this model,and a network structure containing only encoders is adopted,and the modules such as void convolution,deep separable convolution and two-branch structure are combined.Experiments on Cam Vid dataset and Cityscapes dataset demonstrate that the proposed lightweight semantic segmentation model can achieve semantic segmentation in real time with guaranteed accuracy.(2)Semantic segmentation of unmanned aerial vehicles visual scenes is achieved.At present,there are few data sets for semantic segmentation of unmanned aerial vehicles visual scenes,so this thesis produces a high-resolution unmanned aerial vehicles visual image data set according to the principle of large intra-class gap and small inter-class gap.Based on this data set,an improved algorithm based on UNet is proposed in this thesis.The most important feature of the improved algorithm is that the convolution block used for feature extraction is replaced with the residual module to deepen the network structure.While preventing the gradient from disappearing,the network can learn deeper features,which is conducive to improving the accuracy of image segmentation.(3)Achieved target detection for unmanned ground vehicles and unmanned aerial vehicles vision scenes.In view of the error detection and missed detection in target detection,this thesis improves the YOLO V3 algorithm on the basis of clustering to generate the optimal anchor point frame,reconstruct the prediction box filtering algorithm,etc.Compared with the original algorithm,the improved algorithm can not only detect the target object in the simple environment,but also realize the target detection in the complex environment.Moreover,the improved algorithm is more sensitive to the small target object and has a lower rate of missed detection and false detection than the original algorithm.(4)Achieved depth estimation and target distance measurement for unmanned vehicle vision scenes.This thesis innovatively proposes to measure target range based on depth estimation.Based on the depth of the unmanned ground vehicles visual estimation and target distance measurement in the study,the first depth estimation algorithm was proposed based on unsupervised learning,through the network depth estimation output depth map,and then combined with the unmanned vehicle visual target detection in the scene test box,clustering method is used to select the most can represent the depth of the target object value,then the relationship between the depth value and the actual distance to measure the distance.The depth estimation network in this thesis uses binocular images for unsupervised training without the need to input the true depth information of images.To sum up,based on the research of computer vision technology,this thesis explores the application and optimization of the scene perception and understanding algorithm based on convolutional neural networks in the visual scenes of unmanned ground vehicles and unmanned aerial vehicles,providing a new idea for solving the multi-task requirements of unmanned autonomous systems.