Research On Semantic Segmentation For Specific Categories In Low Speed Automatic Driving Scenarios

Semantic segmentation of image is one of the cornerstone problems in the field of computer vision and plays a important role in the field of automatic driving.The purpose of semantic segmentation is to distinguish pixels in an image by category.In an automatic driving task,the vehicle’s perception and understanding of the road environment is important.Subsequent decision-making and planning tasks all depend on the vehicle’s perception and judgment of the surrounding environment.Although semantic segmentation is widely used in automatic driving,there are still some problems.For example,in the low-speed automatic driving scenario,domestic road conditions are complicated and there are a variety of small obstacles on the road surface,which are often of different types,leading to greater difficulty in model learning.Moreover,the categories focused on by the model often account for a relatively small proportion in the whole picture,so specific data enhancement work is urgently needed.In addition,the traditional semantic segmentation model is difficult to learn the edge features of small obstacles.In order to achieve better segmentation effect in the actual environment perception problem,this paper proposes an improvement of relevant technology of semantic segmentation for specific categories in the automatic driving scene.The main research contents of this paper are as follows:(1)In view of the problems that small and medium-sized obstacles in real domestic traffic scenes are complicated in category and number,and the pixel ratio in the collected images is too low,which makes it difficult to train the model.Based on the original data set,this paper proposes a kind of adaptive data enhancement algorithm for real scenes,which can enrich features according to the original data set.Combining the actual scene of the original image to generate new data and providing two common paste strategies can ensure the spatial semantic consistency and geometric size consistency of the enhanced data set.After data enhancement,this paper uses HRNet as the backbone network,converges with Lovasz loss and cross entropy loss,and carries out three classifications for the real scene data set,among which the segmentation of small obstacles is emphasized.Experimental results show that through the proposed real-scenario-like adaptive data enhancement algorithm,the index of the model has been improved by the enhanced data set,reaching 84.23% Io U,which is better than some traditional data enhancement algorithms,which reflects the effectiveness and superiority of the method.(2)Next,in view of the difficulty in learning the edges of small and medium-sized obstacles in real scenes,this paper proposes an auxiliary loss function for contours,which enables the model to learn a specific category of rough contours through the method of xor and pooling at the beginning,thus reducing the difficulty of the model to learn fine contours at the beginning,and reconstructing the framework.By binding the number of iterations of the model with the roughness of the contour,the model gradually learns and strengthens the edge feature information in the number of iterations.(3)In order to solve the problem of low differentiation between small obstacles and other categories,an improved energy probability density model is proposed based on PEBAL(Pixel-wise Energy-biased Abstention Learning)anomaly detection.By improving the original method,the loss part is only targeted at the classes that are concerned in the training focus,instead of being treated as new class to detect anomalies,and reduces the amount of computation.This method first defines the energy value of the probability value of the predicted object,and expands the distance between the concerned class and the common class to make the model more distinguishable from the class of small obstacles.The experiment shows that the two loss function methods can reach 84.64% and84.73% respectively in the relevant indexes.Combined with the experimental visualization results,it is concluded that the loss function proposed in this paper is superior to some losses of the traditional semantic segmentation method.In addition,the loss function method proposed in this paper is easy to expand and use.In order to verify the universality of the method,different backbone network experiments were conducted on Cityscapes data set under the same experimental environment and conditions.The results show that the effect of the loss function proposed in this paper is better than that of the same Focal loss and Tversky loss.It has certain practicability in practical tasks.