Virtual simulation technology can test the performance of automatic driving algorithm in a virtual environment at a very low cost.In order to further ensure the safety of autonomous vehicle,a multi task learning model based on convolutional neural network is designed and implemented,which is used to extract image features and output point cloud data.Then,the point cloud data is pro-cessed into a three channel aerial view containing height,intensity and density.The prediction results of the model in the aerial view are mapped back to 3D space to obtain 3D target detection results.Finally,the results are rendered at the appropriate location accord-ing to the semantic segmentation results,so as to achieve 3D reconstruction of the real driving scene.The results show that the pro-posed model has tended to be stable after 50 iterations of learning,and the average accuracy of the model reaches about 96%.The ac-curacy of the three models is 48.13%,62.43%,75.69%,and the frame rate is 25,which is better than the other three models.The method proposed in this study can realize 3D reconstruction model with complete structure and rich details in complex cases,and can model a variety of objects well.
维普
万方数据