| Sensor sensing technology that accurately identifies traffic scenes enhances the understanding and rapid response to traffic conditions and is one of the key elements to ensuring the safe driving of intelligent driving vehicles.However,due to the variability of traffic participants,lighting and weather conditions,as well as road scenes,there are still certain challenges in accurately identifying and constructing complex traffic scene recognition models.Based on this,in order to improve the performance and accuracy of object recognition and classification in traffic scene images and enrich the understanding methods of traffic scenes,a deep learning model that can recognize and classify traffic scenes more accurately is constructed and verified,and a deep learning method based on LSTM scene understanding is proposed.The main research contents and work of this article are as follows:(1)Aiming at the current low accuracy of traffic scene classification and recognition,different deep neural network models are optimized and compared for analysis.The results indicate that SE＿ResNeXt101 can achieve higher accuracy in the recognition process by learning fewer data sets compared with other CNN(Convolutional Neural Networks)models because of its unique channel attention mechanism and covariance network structure.The SE＿ResNeXt101 model is further compared with another deep neural network model based on the improved Swin-Transformer structure,and it is learned that the Swin-Transformer model has higher accuracy(about 8%),which can be considered the best deep learning model on the given test set.In addition,the TResNetl multi-label recognition model is deeply optimized to more accurately discriminate between vehicles and pedestrians in images by analyzing the generated visual heat maps.(2)To improve the limitations of target detection methods in the field of intelligent transportation,a method that can generate a semantic understanding of traffic scenes based on image description was proposed.This method is based on the understanding dataset constructed by La RA traffic scenes,uses image description generation technology in traffic scene understanding tasks,and creates a semantic understanding encoder and decoder model based on VGG19 and LSTM to provide driving decisions generated by the deep learning model.Based on this,combining the test images in the standard dataset Flickr30 k,this method is compared with other scene understanding methods(NIC、ATT-NIC)in terms of the accuracy of generating evaluation semantics.It is known that this traffic scene understanding method has high accuracy and enriches the functionality of creating scene semantic interpretations.(3)In order to verify the effectiveness and feasibility of the optimized model for traffic scene classification and recognition,Swin-Transformer,SE-ResNeXt101,and ResNet-50 models were deployed on the mobile terminal to collect actual traffic scenes for comparative analysis of experimental data.The results show that,under the same dataset training,the Swin-Transformer model is about 6% higher in recognition accuracy,this also verifies the effectiveness of the actual scene application of the model.To summarize,the traffic scene classification method and scene understanding methods proposed in this paper are anticipated to furnish a suitable illustration for unmanned environmental awareness technology. |
