Research On Semantic Segmentation Algorithm For Autonomous Driving Environment Perceptio

The development of autonomous vehicle technology has gradually improved the traffic safety and road congestion issues in today’s society.As one of the most important links in autonomous visual perception systems,image semantic segmentation algorithm research is of great significance for the development of autonomous driving technology.At present,among mainstream image semantic segmentation algorithms,there is a common problem that the computational burden of the calculation process is too large due to the overly complex algorithm structure.When used in unmanned driving scenes,it is easy to encounter problems such as poor segmentation accuracy and positioning accuracy.Therefore,in response to the above issues and combined with the characteristics of unmanned road scenes,a lightweight semantic segmentation algorithm network based on Attention Mechanism and Deep Separable Convolution is designed.At the same time,an Adam Optimization Algorithm(Adam-GC)based on Gradient Compression(GC)is proposed to improve the segmentation speed,accuracy,and generalization ability and stability of the network model.The main work is as follows:After in-depth research on the principles of image semantic segmentation algorithms,a lightweight feature extraction network was constructed.On the basis of the U-Net algorithm structure,standard convolutions were replaced with deep separable convolutions,and attention mechanisms were introduced to achieve weight information learning of feature channel dimensions.At the same time,gradient compression was used to train and optimize the algorithm during model training to improve the training speed,detection accuracy,and generalization ability of the network model.The Cityscapes dataset is extracted and preprocessed to make its data format match the algorithm model proposed in this paper.Then,the data enhancement methods commonly used in the image semantic segmentation field and the advantages and disadvantages of loss function in the road scene perception task are analyzed.Mosaic data enhancement method is used to enhance the training data,At the same time,the uniform intersection over Union(MIo U)loss function is selected to train the algorithm model proposed in this paper.To verify the effectiveness of the algorithm model proposed in this article,the Cityscapes dataset was used to train the original algorithm at each stage with the algorithm model proposed in this article,and various performance indicators were compared and analyzed.At the same time,the algorithm model proposed in this article was further compared and validated with other mainstream semantic segmentation algorithms.Finally,the generalization ability of the proposed method was verified by setting up generalization experiments.The experimental results of effectiveness verification and comparative verification show that the overall segmentation result of the improved algorithm network can reach78.02% MIo U on the Cityscapes validation set,which is 5.9% higher than the original algorithm.The segmentation effect on important categories of unmanned driving scenes such as pedestrians,vehicles,and roads is relatively excellent,and its category can be accurately identified.The segmentation edges are also smooth and accurate,Better than basic algorithm networks and other mainstream algorithm networks.Meanwhile,compared to the original U-Net,the improved algorithm can achieve a segmentation speed of 0.14 seconds per image,increasing the segmentation speed by 12.5%,meeting the requirements of lightweight algorithms.In the generalization experiment,it can reach78.24% and 79.12% MIo U,and the segmentation speed on the Cityscapes validation set can reach 0.07 seconds.In addition to meeting the requirements of stability,accuracy,generalization,and lightness,it has special significance for the development of image semantic segmentation technology.Therefore,the lightweight image semantic segmentation algorithm network improvement method proposed in this article can not only effectively reduce the computational burden of the autonomous vehicle perception system,but also play a positive role in improving the performance of its perception system.