交通事故现场语义描述分析  

Semantic Description Analysis of Traffic Accident Scene

作　　者：杨俊成[1,2] 李淑霞 YANG Juncheng;LI Shuxia(School of Electronic Information Engineering,Henan Polytechnic Institute,Nanyang Henan 473000,China;School of Computing,Wuhan University,Wuhan Hubei 430072,China)

机构地区：[1]河南工业职业技术学院电子信息工程学院,河南南阳473000 [2]武汉大学计算机学院,湖北武汉430072

出　　处：《信息与电脑》2023年第5期73-76,共4页Information & Computer

基　　金：河南省科技厅科技攻关项目(项目编号:222102210203);河南省教育厅高等学校重点科研项目(项目编号:22B520009);全国高等院校计算机基础教育研究会纵向课题(项目编号:2022-AFCEC-288、2022-AFCEC-295)。

摘　　要：交通事故现场的情景描述对快速处理交通事故具有重要作用,而如何准确快速地描述事故现场是近年来研究的热点和难点。首先,依据交通事故现场的实际情况,利用生成对抗网络(Generative Adversarial Network,GAN)生成新的数据样本,并对图片和图像进行平移、翻转、旋转以及缩放,从而解决交通事故数据样本少的问题。其次,结合Vatic等标注工具自动生成和视觉内容具有极高关联性的标注语句,大大提高样本的质量。再次,利用改进的卷积神经网络(Convolutional Neural Network,CNN)模型深度挖掘数据的多尺度特征,引入多维注意力模型,融合视觉信息,构建基于多维自注意力机制的交通事故描述判别器,充分利用门循环单元(Gate Recurrent Unit,GRU)网络的记忆特点来生成对应的文本描述,实现对交通事故的快速识别和事故场景描述。最后,在Flickr8K和MS COCO数据集上测试,模型都取得较好的效果。The scenario description of a traffic accident scene has an important role in the rapid processing of traffic accidents,and how to describe the accident scene accurately and quickly is a hot and difficult research point in recent years.Firstly,based on the actual situation of the traffic accident scene,we use the Generative Adversarial Network(GAN)to generate new data samples and perform image translation,flip,rotation,and scale to solve the problem of few traffic accident data samples.Secondly,combine with annotation tools such as Vatic to automatically generate annotation statements with high correlation with visual content,which greatly improves the quality of samples.Thirdly,use the improved Convolutional Neural Network(CNN)model to deeply mine the data.We also use the improved CNN model to deeply explore the multi-scale features of the data,introduce the multidimensional attention model,fuse the visual information,build a traffic accident description discriminator based on the multidimensional self-attention mechanism,and make full use of the memory characteristics of the Gate Recurrent Unit(GRU)network to generate the corresponding text descriptions to achieve rapid recognition of traffic accidents and accident scene descriptions.Finally,the model is tested on both Flickr8K and MS COCO datasets,and the model achieves good results.

关 键 词：交通事故 生成对抗网络(GAN) 注意力 图片描述 

分 类 号：TP301.6[自动化与计算机技术—计算机系统结构]