| Image segmentation is the cornerstone technology of image understanding,which plays an important role in the applications of autopilot,UAV and wearable devices.Especially in the field of automatic driving,the technology of road scene image segmentation is very important.Road scene segmentation is to classify every pixel in the visual image,and segment the objects including road,vehicle and pedestrian to get a semantic label image which represents the driving scene After receiving the semantic label image,the computer automatically makes the driving decision according to the segmentation result on the label image.There are many methods to realize image segmentation,including traditional manual feature method,machine learning method and depth learning method.In recent years,the Convolutional Neural Network(CNN)has made breakthroughs in image,speech and text recognition.The deep convolutional neural network is more widely used because of its high accuracy.The advantage of convolutional neural network is that it can automatically learn the hierarchical feature representation by performing layer by layer feature transformation on the original signal,which is more beneficial to feature visualization.At present,the performance of various depth convolutional neural network in the field of image segmentation has exceeded the traditional image segmentation algorithm.However,convolutional neural network based image segmentation algorithms usually do not obtain the context information and relevance between objects,so it is difficult to achieve a good recognition effect for some objects in complex natural environment that is not obvious or complete.To solve this problem,this thesis proposes an improved multi-region attention algorithm,which can capture the contextual information between different regions in spatial dimension and pay attention to the relevace between objects.The background information of the object,especially the long distance information,is used to mark the object which is not obvious.At the same time,this algorithm is embedded into the advanced image segmentation model to improve the segmentation effect.The proposed method is evaluated on several open data sets and compared with several advanced segmentation algorithms.The experimental results show that the proposed method is superior to the Baseline Model It can effectively identify the obscure and incomplete objects which are difficult to be recognized in traditional image segmentation algorithms.Finally,according to the demand of logistics road scene segmentation system,an intelligent driving assistance system is designed and implemented,and the Algorithm model proposed in this thesis is applied to the system to analyze the road scene information Ensure the safety of drivers in the process of logistics transportation.At the same time,according to the particularity of the logistics scene,this thesis designs and realizes an android intelligent logistics platform application,and designs and realizes each module in the system in detail,this application can bring the traditional logistics link to the online operation In order to improve the efficiency of logistics transportation,the intelligent assistant driving system is embedded into the system as the core module,which makes the whole logistics transportation link safe and efficient. |
