Terrain Recognition Based On Deep Learning

Terrain recognition based on vision is an important research direction in the field of computer vision.It is widely used in the field of robot navigation and autonomous driving.It is a key issue for autonomous navigation of robot.Terrain recognition technology can make the robot fully understand the surrounding environment and realize the accurate prediction of the terrain,so as to adjust the control decision and path planning in time.Because the outdoor terrain is interlaced and complex,the image-level terrain recognition in the face of a variety of terrain co-existence in the case of the recognition effect is not ideal,and pixel-level terrain recognition can identify a variety of terrain at the same time.Therefore,the research work of this paper is carried out in both the image-level terrain recognition and the pixel-level terrain recognition,as follows:Firstly,the significance of research and the current situation of research at home and abroad are summarized from the aspects of image-level terrain recognition and pixel-level terrain recognition.The shortcomings of the existing image-level terrain recognition algorithm and pixel-level terrain recognition algorithm are summarized,and the development of deep neural network and algorithm principle are introduced in detail.Secondly,in view of the problem that the terrain recognition effect is not ideal due to the influence of the complex and environmental light of the wild terrain,the deep residual texture network(DrtNet)is proposed to build a texture detail layer in the residual convolution network to form an end-to-end learning network.The advantage of DrtNet is that it can extract not only the spatial geometry of the terrain image,but also the texture detail feature of the terrain image,which makes DrtNet have a better classification effect.In addition,due to the relatively few public image data sets in the terrain recognition direction,the terrain image under various conditions was collected with the help of camera,mobile phone camera and other equipment,and the terrain data set SDU_Terrain16 containing 16 kinds of typical terrain images was established.Thirdly,in view of the problem that the division effect of complex images and shaped objects in semantic image segmentation is not ideal,the multi-scale segmentation network(MSNet)is proposed,and ResNet is used as the basic network to extract and top-sample the output of ResNet,so that MSNet can not only extract the global characteristics of the image,but also integrate the context feature information;Fully optimize network parameters.Finally,summarize all the work in this paper,point out the unresolved problems in terrain recognition and look forward to future research directions.