Research On Multi-Objective Segmentation Algorithm In Environment Perception For Autonomous Driving

Environment perception is one of the important problems to be solved in autonomous driving,and vision-based environment perception technology has been widely used in the field of autonomous driving,which mainly applies target detection and segmentation algorithms to perceive and understand the scene information around the vehicle.Traditional target detection and segmentation algorithms require manually designed features,resulting in low accuracy and complicated steps.The deep learning-based image detection and segmentation algorithm can perfectly solve the above problems in the field of autonomous driving environment perception.Therefore,this paper studies multi-target detection and segmentation in complex traffic scenes based on semantic segmentation algorithm and instance segmentation algorithm to improve the accuracy of environment perception technology in autonomous driving,and the main research contents are as follows:(1)A lightweight multi-objective semantic segmentation algorithm M2-PSPNet was proposed for regular road autonomous driving environment perception.The algorithm was based on the Pyramid Scene Parsing Network(PSPNet)semantic segmentation model,replaced the backbone network Res Net with the lighter Mobile Net V2 and reconstructed it.The number of model parameters can be reduced while expanding the field of perception to capture more comprehensive contextual information.The Pyramid Pooling Module(PPM)was improved by using a combination of Bilinear interpolation and transposed convolution for upsampling,so that the output feature map could retain more spatial feature information.Finally,the Focal Loss function was introduced as the classification loss function of the model to effectively optimize the problem of poor target recognition accuracy caused by the imbalance of samples in the dataset.The experimental results show that the model maintains high segmentation accuracy with reduced number of parameters.The MIo U on the autonomous driving dataset City Scapes reaches 73.58% and the MPA reaches 86.23%,which are 3.43% and 1.92% higher than the original PSPNet model,respectively.The detection speed on the experimental platform is 0.012 s,which is nearly double compared to the original PSPNet model.(2)An improved Mask R-CNN based multi-target detection and instance segmentation model for autonomous driving environment perception was proposed to improve the Mask R-CNN algorithm for more complex traffic scenarios.Firstly,the backbone network Res Net was replaced by Res Ne Xt network with group convolution to further improve the feature extraction capability of the model,and a bottom-up path enhancement strategy was added to the Feature Pyramid Network(FPN)to achieve feature fusion.The Effificient Channel Attention(ECA)module was also added to the backbone feature extraction network to optimize the high level low resolution semantic information graph.Finally,the Bounding Box loss function smooth L1 Loss was replaced with CIo U Loss to speed up the model convergence and minimize the network error.The experimental results show that the improved Mask R-CNN algorithm achieves 62.62% m AP for target detection and 57.58%m AP for segmentation on the autonomous driving dataset City Scapes,which are 4.73% and3.96% better than the original Mask R-CNN algorithm,respectively.