Terrain Recognition Algorithm Based On Multi-Feature Fusion And Semantic Segmentation

The research field of terrain recognition based on vision is an important research content of computer vision.This research technology has been widely used in the field of robot navigation and autonomous driving,and it is also the key to enable robots to run smoothly and effectively avoid roadblocks.The technology enables the robot to accurately capture the surrounding environment,and then accurately predict the terrain,so that the robot can timely carry out path planning and attitude selection.The research work of this thesis is as follows:1、Aiming at the problems of too few public terrain recognition datasets or uneven classification,this thesis performs data enhancement operations on the Deep Sat SAT-6dataset.Since the texture features of the terrain image are distributed throughout the image and the natural light of the environment has a great influence on the image,the main step of data enhancement in the thesis is to rotate at any angle first,and then change the color of the image to improve the data set.The perfect data set provides the training basis for the following network model.2、To solve the problem that the recognition effect is not ideal due to the large difference between terrain data set and actual environment,a deep residual color texture network(Drct Net)was proposed,which added a texture detail layer after the residual convolution structure to extract texture features.And a color feature layer is parallel with the residual convolution network to extract color features.The advantage of Drct Net is that it can simultaneously extract the overall spatial feature,color detail feature and texture detail feature of terrain image,which makes Drct Net have better classification effect.3、Due to the limited receptive field size and lack of cross-channel interactions of the basic networks,they may not be suitable for target detection,image segmentation and other fields.This means that to improve the performance of a given computer vision task,the underlying framework of Res Net needs to be modified to make it more effective for the specific task.Compared with existing Res Net variants,Res Ne St improves the network capability without requiring additional computation and premise.And Res Ne St can serve as a skeleton for other tasks,thereby improving performance across multiple tasks simultaneously.Therefore,the thesis improves the basic network of the deep residual texture network Drct Net,introduces a modular distraction module,and proposes a deep residual color texture attention mechanism network(Drct Ne St).4、Aiming at the difficulty of identifying the complex environment of multiclassified terrain,this thesis proposes a classification method based on Improved HighResolution Object Contextural Representative Network.On the basis of not greatly increasing the amount of computation,this thesis adds the aggregated object contextual representations(OCR)on the basis of improving the high-resolution network(HighResolution Network,HRNet V2)to improve the network structure.The parallelization of feature maps of different resolutions is implemented to preserve the information and the pixel representation is enhanced by the upper and lower sub-association,which enhances the segmentation effect.Next,the model is refined by adding a model-independent postprocessing scheme to improve the boundary quality of the segmentation results generated by any existing segmentation model,replacing the original unreliable boundary pixel predictions with predictions of interior pixels.