Research On Image Semantic Segmentation And Semantic SLAM Of Unmanned Vehicle Scene

As an important information of image understanding,image semantic information plays an important role in scene understanding and positioning composition of unmanned vehicle.Image semantic segmentation is a common technology to extract image semantic information.With the development of convolutional neural network,the method based on deep learning has surpassed the traditional method and become the mainstream method of image semantic segmentation.In the process of driving,the scene of no light at night has more security risks than the scene of day.Different from the images collected by the visible light imaging equipment in the daytime,the infrared images obtained by the infrared thermal imager based on the temperature distribution at night have the characteristics of less texture information,more noise and low contrast,so it is more difficult to segment.The method of deep learning is used to realize the image semantic segmentation under the day and night scenes of the unmanned vehicle,obtain the semantic information of the surrounding environment of the unmanned vehicle,expand the scene understanding ability of the unmanned vehicle,and effectively help the unmanned vehicle to make driving decisions or guide the unmanned vehicle to realize its own positioning and environment composition.With the development of machine vision,visual simultaneous localization and mapping(VSLAM)technology gradually plays an important role in the field of unmanned vehicle navigation.At present,VSLAM technology is mostly based on the low-level features or pixel points of the image,which can not meet the long-term invariant and static needs of the features,while a large number of redundant features increase the complexity of map storage,so it is very unfavorable to maintain a long-term map.Using the image semantic information extracted by deep learning toguide the selection of VSLAM features,the VSLAM system tends to select some static and long-term invariant features,which is conducive to the long-term positioning of unmanned vehicles.This paper mainly studies the image semantic segmentation and semantic VSLAM based on the unmanned vehicle driving scene.The research content mainly includes two parts: one is the semantic segmentation of UAV image based on the improved deeplabv3 +;the other is the semantic slam technology based on the combination of the improved deeplabv3 + network and orb-slam2 based on information theory.The main innovations are as follows:1.An improved deeplabv3+ network is proposed to realize the semantic segmentation of the image obtained by the unmanned vehicle.For the driverless scene,a dense connected ASPP module is introduced,which enables the multi-scale features generated by the network to be correctly encoded while covering a larger scale range,and does not cause additional computing time loss.At the same time,in order to recover more spatial information lost in the process of downsampling,the multi-layer low-level feature map of the encode module of the original deeplabv3+ network is integrated into the decode module,which effectively improves the segmentation effect.2.A semantic slam technology based on information theory is proposed,which combines deeplabv3+ network with ORB-SLAM2.In this method,the uncertainty of the improved deeplabv3+ network classification results is introduced into the feature selection pipeline.By adding the improved deeplabv3+ network classification entropy into the state joint entropy,the features with large mutual information between the current state entropy and the state joint entropy are selected,which significantly reduces the uncertainty of vehicle state.At the same time,the detected features are required to belong to the static object,with a high degree of credibility.At the same time of ensuring the accuracy of the map,the size of the map is greatly reduced,which is conducive to long-term positioning and mapping.