Research On Semantic Segmentation Of Urban Traffic Image Based On Transfer Learning

As an important part of urban management,intelligent traffic system(ITS)requires more capability of scene information recognition.Semantic Segmentation,the basic technology of image understanding,can provide more effective tools and solutions for ITS.The assumption of many machine learning algorithms is that the training and test data must be independently identically distribution.However,in many urban traffic scenarios,this assumption may not hold,resulting in significant decrease in accuracy when model trained on training data is transferred to the actual scenario.Therefore,based on the analysis of the three migration situations in the semantic segmentation of urban road traffic images,we propose a semantic segmentation model of urban traffic image based on transfer learning to reduce the decrease of accuracy.With reference to the structure of public datasets,the high-altitude monitoring dataset was constructed through manual image annotation and dataset division based on the monitoring images collected by surveillance camera.The virtual images are acquired through the communication between the monitoring game and the graphics library,the semi-automatic annotation is implemented based on the custom scripts,and GTA virtual dataset are constructed.For three migration situations,three migration learning datasets are constructed.Algorithms such as PCA jittering and random image cropping and patching are used to augment the above dataset.The existing semantic segmentation model is improved from the aspects of attention mechanism,auxiliary tasks,context mechanism,training algorithm.And a semantic segmentation network model called Road Net is designed: the identity residual network is improved based on Attention mechanism.Auxiliary task learning algorithm is proposed based on multi-task learning algorithm.Edge detection,image reconstruction and saliency detection are selected as auxiliary task for semantic segmentation.Cascade and global pooling is introduced in atrous space pyramid pooling to design cascade atrous space pyramid pooling.Prune-Initialization-Prune training algorithm is designed to maximize the potential of models.The experimental results on the VOC2012,Cityscapes,and high-altitude monitoring datasets show that the mean intersection over union of Road Net are 86.2%,84.5%,and51.8%,outperforming the existing model by 3.3%,3.7%,and 2.7%.The transfer learning algorithms for semantic segmentation is improved from the aspects of domain discrepancy measurement,context semantic distribution and adversarial learning algorithm.Semantic segmentation based on transfer learning called Transfer Road Net is designed.: MMD is selected as domain discrepancy measurement;semantic distribution and context semantic distribution module are designed to force model to learn domain-invariant features;domain discriminator and adversarial learning algorithm are designed based on generative adversarial network.The experimental results on the three transfer datasets show that the mean interaction over union of Trans Road Net are62.7%,30.6%,and 35.8%,outperforming the model without transfer by 4.1%,24.4%,and 12.7%,also outperforming common transfer learning algorithms by 1.9%,3.7% and 4.4%.